Agent sandboxes, compared
Every company we found that sells sandboxes or cloud computers for AI agents, Orble included.
We took each price from the company's pricing page and normalized it to 2 vCPU and 4 GiB where we could. We checked them on 14 November 2026. To correct a row, email info@orble.tech.
| 2 vCPU, 4 GiB | When idle | Isolation | Notes | ||||
|---|---|---|---|---|---|---|---|
| Core | Core | Monthly plans: Developer $39/mo (1 environment, 1 shared vCPU, 3 GiB RAM, 10 GB storage)Monthly plans: Developer $39/mo (1 environment, 1 shared vCPU, 3 GiB RAM, 10 GB storage); Pro $79/mo (3 environments, each 2 shared vCPU, 8 GiB, 50 GB); annual saves 25%; Enterprise custom; 7-day trial (512 MiB, 5 GB, 25% of one core); no 2 vCPU/4 GiB size | Included in the monthly plan (flat fee) | Sleeps after the client disconnects and stays idle 30 min (Developer) or 2 h (Pro)Sleeps after the client disconnects and stays idle 30 min (Developer) or 2 h (Pro); sessions and files persist; external connections may need to reconnect after wake | VM | Mar 2026 | Persistent interactive terminals for agents over MCP: signals when a program waits for terminal inputPersistent interactive terminals for agents over MCP: signals when a program waits for terminal input, keeps REPL/database sessions between tool calls; full root; agend CLI configures MCP clients (Claude Code etc.); bring your own agent and model; listed in the official MCP Registry |
| Core | Core | $0.006 | $0 compute; disk $0.09/GB-month ($0.36/mo for 4 GB) | Auto-sleep is opt-in: after idle_timeout_seconds with no trafficAuto-sleep is opt-in: after idle_timeout_seconds with no traffic, memory is saved to disk, CPU/RAM released and only disk bills; next request wakes it; stop/start/resize keep disk (VM cold-boots, memory lost); sandboxes sleep if the prepaid balance runs out; daily backups kept a week | gVisor | Oct 2026 | Hosts agent harnesses (Hermes, OpenClaw, Claude Code, Codex, OpenCode, Grok, Pi) or your own Docker imageHosts agent harnesses (Hermes, OpenClaw, Claude Code, Codex, OpenCode, Grok, Pi) or your own Docker image; VM type runs Docker, k3s, Android emulator (/dev/kvm), XFCE desktop with Firefox; SSH and exec API; authenticated per-port URLs and custom domains; regions us, eu (France), ap-south (Mumbai), VM-only outside US; egress firewalled to public internet only; sizes 2/4 vCPU up to 32 GB RAM, disk 2-100 GB; SOC 2 Type I; claims 94% lower cost than most providers |
| Core | Core | $0.315 | $0 compute when stopped; cold storage $0.000027/GB-hour | Computers run until stopped; stop and start through API or CLI (disk kept as coldComputers run until stopped; stop and start through API or CLI (disk kept as cold storage); snapshots save a computer's state and restore into a new computer; no auto-stop or idle timeout stated | VM | Mar 2026 | Persistent Ubuntu computers with root, SSH, browser access and VNC desktopPersistent Ubuntu computers with root, SSH, browser access and VNC desktop; size presets 2, 4 or 8 GiB RAM (custom 2-64 GiB RAM, 5-250 GiB storage); exec API, file operations, mounts; public or private published ports; snapshots and restore; resize; sharing and share links; macOS computers (QEMU/KVM, 4 GiB default); ~1 Gb/s network and ~3 GB/s NVMe claims; CLI (`computer`), TypeScript SDK, REST API with OpenAPI spec; host and guest image repos on GitHub (no license, 2 stars each) |
| Core | Core | $0.00025/s of compute ($0.90/h), sandbox CPU/RAM size not stated$0.00025/s of compute ($0.90/h), sandbox CPU/RAM size not stated; $0.0005/MB for output files; prepaid credits $5 to $1,000; no subscription; $10 free credits with Google sign-in | — | Without a session_id each execution gets a one-shot sandbox that is destroyed afterwardsWithout a session_id each execution gets a one-shot sandbox that is destroyed afterwards; sessions keep /workspace files and installed packages across executions until deleted; session idle expiry not stated | — | Feb 2026 | REST API and Python SDK (agentsandbox-sdk, MIT); Python and Bash execution; file upload/download and injectionREST API and Python SDK (agentsandbox-sdk, MIT); Python and Bash execution; file upload/download and injection; environment variables; persistent sessions; OpenClaw plugin and a skill.md for Claude Code, Codex, Cursor, Windsurf, Cline |
| Core | Core | Per request, paid in USDC on Base via HTTP 402 (x402)Per request, paid in USDC on Base via HTTP 402 (x402), no account: $0.01 at default limits (5 s timeout, 2 GiB memory ceiling, 1 MiB output, 64 MiB work dir); tiers standard $0.02, elevated $0.03, high $0.05, max $0.10 (up to 60 s, 4 GiB, 8 MiB output, 256 MiB work dir); POST /quote prices a limits object; one free 2-second run per wallet per UTC day | $0 (nothing persists between requests) | Per-request run to completion: timeout 5 s by default, up to 60 sPer-request run to completion: timeout 5 s by default, up to 60 s; no persistent sandbox (separate paid /storage endpoint keeps bytes for a retention window) | — | — | Accountless pay-per-use code execution for agents: python3 (standard library plus numpy 2.2 and scipy 1.15) and nodeAccountless pay-per-use code execution for agents: python3 (standard library plus numpy 2.2 and scipy 1.15) and node (standard library only); no network inside the sandbox; extra files and stdin can be sent; MCP endpoint with a paid execute tool; the same host sells many other x402 utilities (storage, claim, attest, watch, address/URL checks, disposable inbox) and a directory of x402 services |
| Core | Core | $0.027 | Stopped: 5% of the running price ($0.99/month for small, disk kept)Stopped: 5% of the running price ($0.99/month for small, disk kept); with memory kept 10% ($1.97/month for small) | Runs until stopped unless given a time limit (ttlSeconds, at most 24 hRuns until stopped unless given a time limit (ttlSeconds, at most 24 h, which the API or the agent inside can keep moving); no default limit; stop keeps the disk, or the disk and memory with keepMemory (start then restores every process); stopped 6 h -> archived (disk moved to storage, same price; memory-kept sandboxes are not archived); schedules can start a stopped sandbox, run a command and stop it again; when credits run out, running sandboxes are stopped with disks kept | VM | Oct 2026 | TypeScript and Python SDKs (MIT), CLI (npx agental), OpenAI Agents SDK and GitHub Actions integrationsTypeScript and Python SDKs (MIT), CLI (npx agental), OpenAI Agents SDK and GitHub Actions integrations, Claude Managed Agents guide; fork (disk, or with memory on microVM hosts), snapshots and restore, templates; secrets kept out of the sandbox (stand-in values swapped on outbound requests); outbound rules held outside the VM; SSH, web terminal, desktop with screenshots and input for computer use, Chrome preinstalled; up to 5 public ports; an email address per sandbox; cron schedules, webhooks, OpenTelemetry export, per-minute metrics; regions eu (default; EEA or UK), us and apac where they have servers; create under a minute; GDPR, EU-hosted, DPA; VERTIXER LTD; GitHub agental-io (SDK mirrors, 0 stars) |
| Core | Core | $0.032 | $0 compute while paused; storage $0.0001/GiB-hr | Sandbox timeout 30-3600 s (default 600 s), extendableSandbox timeout 30-3600 s (default 600 s), extendable; pause sleeps the container and keeps the filesystem in R2 with no compute charge; resume restores; destroy removes R2 blocks; max session 30 min Free / 24 h Pro | V8 isolate | — | Cold start under 50 ms (pure JavaScript sandbox, bash interpreter with 40+ POSIX commands, no native Linux)Cold start under 50 ms (pure JavaScript sandbox, bash interpreter with 40+ POSIX commands, no native Linux); runs in Cloudflare Workers, Node.js, Bun or browsers; one API over in-process, Docker, Cloudflare and bare-metal providers; TypeScript SDK @agentrunbox/sdk and REST API; templates base/node/python/rust/go; git, files, processes APIs; says Apache-2.0 open source and self-hostable, but its GitHub link (github.com/codeboltai/codeboltshell) returns 404 |
| Core | Core | $0.1426 | $0 while paused (docs: paused sandbox is not charged); storage free | Sandbox shuts down 5 min after start by default (timeout settableSandbox shuts down 5 min after start by default (timeout settable, set_timeout extends it); max session 12 h Free / 24 h Pro; pause saves filesystem and memory (running processes) and resume restores them, but not past the sandbox's end time; auto-pause after 10 min inactivity (beta); kill deletes permanently | VM | Jul 2025 | Cold start as low as 100 ms; pause/resume with memory; snapshot recovery; MCP server for MCP clientsCold start as low as 100 ms; pause/resume with memory; snapshot recovery; MCP server for MCP clients; desktop sandboxes for computer-use agents; Python and JS/TS SDKs and CLI; custom templates and CPU/RAM; private deployment in your own cloud (AWS, GCP, on-prem); claims SOC2 and GDPR; pitches itself as an E2B alternative |
| Core | Core | Live quotes per provider plan and region, shown as $/machine-hour (no fixed rate card onLive quotes per provider plan and region, shown as $/machine-hour (no fixed rate card on the public pages); the pricing page says 'base retail floor $0.09/hour, 80% markup' for direct catalog pricing, with HelixCore partitioned capacity priced dynamically; machine-payments (MPP/x402) fixed-budget runs: hetzner-shell $0.50 to $25, hetzner-mcp-sandbox $0.50 to $50; compute quote $0.50 per request | Machines are destroyed at expiry; no paused state described | Bounded runtime: you set a runtime limit and maximum charge before startBounded runtime: you set a runtime limit and maximum charge before start, then stop manually or the machine is destroyed automatically at expiry (machine-payable runs 600 to 21,600 s); no pause or suspend described | VM | Aug 2026 | Dashboard, CLI (npm @wundercorp/agentvm, Apache-2.0) and MCP; SSH; optional RealVNC GUI setupDashboard, CLI (npm @wundercorp/agentvm, Apache-2.0) and MCP; SSH; optional RealVNC GUI setup; images Ubuntu, Debian, Fedora, Rocky, Alma, CentOS Stream; preinstalled Codex, Claude Code, Cursor Agent, Gemini CLI, OpenClaw, LokiAgent; agents can pay with Stripe MPP or Tempo sessions; shell runs 1-4 vCPU / 1-16 GB |
| Core | Core | Pay as you go after a $10 minimum prepaid credit (Stripe or x402): Small (1 vCPUPay as you go after a $10 minimum prepaid credit (Stripe or x402): Small (1 vCPU, 1 GB) $0.06/min ($3.60/h), Medium (2 vCPU, 2 GB) $0.12/min ($7.20/h), Large (4 vCPU, 4 GB; Pro/Business only) $0.24/min ($14.40/h); billed per started minute with a 1-minute minimum; queue and provisioning not billed. Plans: Starter $5/mo (100 compute-minutes, then 4c), Pro $15/mo (400, then 3c), Business $39/mo (1,200, then 2.5c). No 2 vCPU / 4 GB size | $0 after teardown (idle time while the VM exists is billed) | Jobs default to a 2-minute lifetime, maximum 15 minutes (900 s)Jobs default to a 2-minute lifetime, maximum 15 minutes (900 s); the VM is destroyed once execution and exports finish; ready time, including idle time, is billed; explicit sessions need DELETE for early cleanup; exported artifacts kept for the plan's retention window (1, 7 or 30 days) | VM | — | HTTP API for bounded jobs: Python, Node and shell in an approved Linux imageHTTP API for bounded jobs: Python, Node and shell in an approved Linux image; guests start offline as a non-root user on a read-only base image; files in and artifacts out up to 8 MB; memory-backed scratch capped at 2/4/8 GB; no custom images, inbound ports, snapshots or persistent volumes; agents can sign up over HTTP and pay with USDC on Base via x402; every paid run needs a spending cap; OpenAPI and llms.txt; paid AGNT plans include 100-600 compute-minutes |
| Core | Core | Monthly, paid per call in USD stablecoins (x402 on Base, Polygon or SolanaMonthly, paid per call in USD stablecoins (x402 on Base, Polygon or Solana, or MPP on Base): vps-small (1 vCPU, 1 GB, 10 GB NVMe) $10.012 first month, renewals $10.00/month; vps-medium (2 vCPU, 2 GB, 25 GB NVMe) $14.612 first month, renewals $14.60/month; live quotes are authoritative; no 2 vCPU / 4 GB plan | N/a (paid by the month; no paused state) | Always on for the paid month; no auto-renewAlways on for the paid month; no auto-renew; renew 1-6 months per call, at most 12 months ahead; an expired server is removed with its data | — | Oct 2026 | Servers for AI agents rented by the month with no account, API key or cardServers for AI agents rented by the month with no account, API key or card; Ubuntu 24.04 with root and public IPv4 in Germany (default) or Finland; ready about 5-10 minutes after payment; REST API with OpenAPI spec and an MCP server (listed in the official MCP registry as dev.shveik/agvps); status, password reset and reboot endpoints; built by shveik.dev |
| Core | Core | Rates not published ('We have not fixed the numbers')Rates not published ('We have not fixed the numbers'); billed on wall-clock session runtime from ready to stop, idle time inside a running session included; Shared tier (default) and Dedicated tier (reserved bare metal, terms per team); model tokens are your own and billed by the provider | $0 after a session stops; idle time inside a running session is billed | Optional idle timeout (the session stops itself)Optional idle timeout (the session stops itself), optional per-session lifetime cap and per-account runtime-minute budget, both with auto-stop; no runtime ceiling; session persistence, resume and fork are on the roadmap, not built | VM | Jul 2026 | Managed control plane for AI coding agents run against your git repositoriesManaged control plane for AI coding agents run against your git repositories; bring your own model token or an EU-hosted model; the token is a placeholder in the guest and substituted on outbound requests; egress reaches only destinations named in the session's configuration; live session event feed stored as OCSF records; runs on bare metal at an EU provider; Dedicated tier gives physical host separation; roadmap: self-service egress allowlist, event export, session persistence/resume/fork, public CLI and agent skill, on-premises, external pen test; DPA not yet available |
| Core | Core | Tech preview; the product page gives no pricePricing page | — | Intermittent agents: each step completes in about a minuteIntermittent agents: each step completes in about a minute, then the agent stops until its next instruction; key-value storage keeps context between steps; stateful components for agents that stay alive for days or weeks are planned | WebAssembly | — | Extends Akamai Functions with WebAssembly isolation and deny-by-default outbound accessExtends Akamai Functions with WebAssembly isolation and deny-by-default outbound access; the developer grants access to specific external services; execution limits enforced at runtime so neither the agent nor its code can raise them; Spin SDK (Spin is a CNCF project); works with any OpenAI-compatible model |
| Core | Core | $0.0714 | Light hibernation: memory + disk (no vCPU)Light hibernation: memory + disk (no vCPU); deep hibernation and snapshots: disk only, billed on memory plus disk size, without the 15 GiB free allowance ($0.00025308/GiB-h Economy outside China) | Active, light hibernation (Flash Edition onlyActive, light hibernation (Flash Edition only; no vCPU charge, memory and disk billed), deep hibernation (memory written to disk; only disk billed), snapshot saved, destroyed; resume 'within seconds'/'millisecond-level'; Flash Edition sandbox stays active up to 86,400 s (24 h) per continuous run; idle auto-hibernate timeout not stated | VM | — | Two editions: Flash Edition (managed developer API, E2B SDK compatible) and Container Edition (Kubernetes API with aTwo editions: Flash Edition (managed developer API, E2B SDK compatible) and Container Edition (Kubernetes API with a Sandbox CRD on ACS, E2B protocol compatible); code, browser, Windows desktop and Android mobile sandboxes; checkpoints while running and batch-cloning tens of thousands of sandboxes from one checkpoint; claims 100,000 sandboxes created per minute and million-scale concurrency, P99 under 1 s for 5,000 concurrent heterogeneous image creations; Flash limits: 300 sandboxes per region, 15 GB temp disk, 10 GB template image, 1-5 Gbit/s; zero-trust access to cloud APIs and LLM keys without injecting secrets; L4/L7 network policies |
| Core | Core | $0.0992 | Paused (beta): lower than running, rate not statedPaused (beta): lower than running, rate not stated; released sessions cost nothing; storage $0.000045/GiB-hour beyond the plan quota | Session auto-released after an idle timeout (LifecyclePolicy default 5 minSession auto-released after an idle timeout (LifecyclePolicy default 5 min, minimum 3) or at max runtime (default 30 min), or never with manual release; keep_alive resets the idle timer; released sessions cannot be recovered; beta pause/resume (allowlist only) preserves all session state at lower cost; context storage persists data across sessions | VM | May 2025 | Four environments: Browser Use (Playwright and Puppeteer compatible, headless or headed)Four environments: Browser Use (Playwright and Puppeteer compatible, headless or headed), Computer Use (Windows and Linux desktops, custom images), Code Space (cloud dev environment with toolchains) and Mobile Use (cloud Android); SDK and MCP integration; built-in Browser Use and Mobile Use agents; process allowlists and blocklists; session replay; endpoints China (Shanghai, default) and Singapore; open-source SDK (Apache-2.0, 1.1k stars) |
| Core | Core | $0.1739 | $0 compute when stopped (billed only for the seconds a rig runs)$0 compute when stopped (billed only for the seconds a rig runs); whether storage ($0.0001134/GiB-hour) is charged while stopped is not stated | Stop keeps a persistent rig's filesystem and environmentStop keeps a persistent rig's filesystem and environment, start resumes it ('stopped' = paused, resumable); temporary rigs are deleted when work is done; VM snapshots save a rig's filesystem as a base image; billed only while running; no auto-stop timeout stated | VM | Feb 2026 | Devbox VMs ('rigs') for coding agents with agent harnesses such as Claude Code, Codex, OpenCode or PiDevbox VMs ('rigs') for coding agents with agent harnesses such as Claude Code, Codex, OpenCode or Pi; runs on Amika's hosted compute or your Mac Mini, desktop or EC2 via a host daemon; private mesh network across rigs; SSH, private HTTPS preview URLs, Cursor and Codex connections; secrets kept unreadable by the agent; start agents from Slack, Linear, Sentry, CLI, TypeScript SDK or HTTP API; macOS rigs; GitHub Actions runners; open source (Apache-2.0, 303 stars); by Fixpoint, Inc. |
| Core | Core | $0.284 | Paused: snapshot storage only, $0.03/GB-monthPaused: snapshot storage only, $0.03/GB-month; an idle running sandbox still pays $0.012/h plus its resident memory | Default timeout 300 s, then delete, or pause if lifecycle on_timeout is 'pause'Default timeout 300 s, then delete, or pause if lifecycle on_timeout is 'pause'; optional pause after N idle minutes; pause keeps memory and running processes (engine test hosts: pause 0.1-0.3 s, resume 0.8-0.9 s; production numbers not published); timeout -1 means no deadline on Pro and Scale; longest timeout 1 h on Free and 24 h on Starter; paused sandboxes kept 5, 50, 200 or 800 by plan; auto-resume coming soon | VM | Nov 2026 | E2B SDK compatible: the stock e2b Python and JavaScript SDKs work with E2B_DOMAIN and E2B_API_KEY (checked against e2bE2B SDK compatible: the stock e2b Python and JavaScript SDKs work with E2B_DOMAIN and E2B_API_KEY (checked against e2b 2.20.3, 2.51.0 and 2.61.0); Anona SDKs (Python 'anona', TypeScript '@anona-labs/compute-sdk') add live view and SSH; CLI and MCP server; fork; templates; desktop; SSH with 10-minute certificates naming one sandbox; credential vault included; Ubuntu 24.04 base; region us-east-1; 5 to 500 running sandboxes by plan; SDK repo anonalabs/anona-compute-sdk (Apache-2.0); Anona Labs Inc., San Francisco, which also sells Anona Memory |
| Core | Core | Pay per platform usage in compute units: 1 CU = 1 GB of memory for 1 hourPay per platform usage in compute units: 1 CU = 1 GB of memory for 1 hour; 1 CPU core per 4,096 MB (memory 128 MB to 32 GB, disk 2x memory); Actor default 4,096 MB = 1 core = 4 CU/h; CU price $0.20 on Free ($5 monthly usage) and Starter ($19/mo), $0.16 on Scale ($199/mo), $0.13 on Business ($999/mo); no 2 vCPU/4 GiB size: 4,096 MB (1 core) = $0.80/h and 8,192 MB (2 cores) = $1.60/h at $0.20/CU; GB treated as GiB | Same as running until the idle timeout (default 900 s) ends the runSame as running until the idle timeout (default 900 s) ends the run; storage billed separately | Idle timeout 900 s by default (0 disables) ends the runIdle timeout 900 s by default (0 disables) ends the run; Actor execution timeout default 0 (no limit); each request capped at 5 min; state persists across container migrations | Container | Dec 2025 | Apify Actor with an MCP endpoint (/mcp: execute, read-file, write-file, list-files), REST /exec and /fs APIsApify Actor with an MCP endpoint (/mcp: execute, read-file, write-file, list-files), REST /exec and /fs APIs, and a browser terminal (ttyd) at /shell; Python 3 and Node.js; Claude Code, Codex CLI and OpenCode preinstalled; setup script; bridges to expose servers running inside the sandbox; 1,791 users and 3,702 runs per the Apify API |
| Core | Core | $0.0020 per POST /execute request, paid via x402 (USDC on Base) or prepaid APIHub credits$0.0020 per POST /execute request, paid via x402 (USDC on Base) or prepaid APIHub credits (minimum top-up $5); 25 free calls per month | $0 (per request) | Stateless per request; default timeout 5 s, maximum 10 s | — | — | JavaScript and TypeScript only; returns result, console logs and execution time; network access blocked; no filesystemJavaScript and TypeScript only; returns result, console logs and execution time; network access blocked; no filesystem; listed on APIHub, an x402 API marketplace for agents, with MCP server and CLI (@apihubio/cli) |
| Core | Core | $0.1656 | $0 compute when a Computer is stopped$0 compute when a Computer is stopped; disk $0.10/GiB-month, snapshots $0.05/GiB-month | Three forms: Computer (persistent; snapshot disk + memory, stop keeps diskThree forms: Computer (persistent; snapshot disk + memory, stop keeps disk, resume later), Sandbox (ephemeral, destroyed when the job ends or on a TTL), Runner (fresh per CI job, torn down after); stopped Computers are billed only for storage | VM | Jan 2026 | Snapshot restore ~22 ms, warm create ~50 ms; Linux (x86-64, arm64Snapshot restore ~22 ms, warm create ~50 ms; Linux (x86-64, arm64; any OCI image), Windows 11 / Server 2022 with full desktop for GUI agents, macOS on Apple silicon, iOS simulator, Android emulator; CLI abctl, Python and TypeScript SDKs, REST API; lifecycle event stream; per-environment network rules; open-source runtime ArcBox Core (GitHub arcboxlabs/arcbox, Apache-2.0 per GitHub, 4,309 stars; site says MIT/Apache-2.0); also a local macOS desktop app |
| Core | Core | $0.19 | $0 compute when paused or stopped; disk storage $0.20/GiB-month active +$0 compute when paused or stopped; disk storage $0.20/GiB-month active + $0.025/GB-month archive | Pauses after 30 min with no process connection (idleTtlSecondsPauses after 30 min with no process connection (idleTtlSeconds, 0 disables) and after at most 24 h per session; pause keeps memory, processes, root filesystem and disks; stop keeps filesystem and disks but drops memory; Free sessions max 30 min; serverless tool calls run one command per microVM, max 5 min | VM | Apr 2026 | Durable S3-backed disks that mount in Archil sandboxes or any Linux machine (bring your own S3/GCS/R2 bucket)Durable S3-backed disks that mount in Archil sandboxes or any Linux machine (bring your own S3/GCS/R2 bucket); fork sandboxes using disk checkpoints and branches; exposed ports with port tokens; supervised services; public and private OCI images; egress firewall with credential brokering; TypeScript (disk) and Python SDKs plus AI SDK tools; regions AWS us-east-1, us-west-2, eu-west-1 and GCP us-central1; 20/250 concurrent sandboxes on Developer/Team; SOC 2 on Team, HIPAA and BYOC on Enterprise |
| Core | Core | Not published: pricing page needs sign-in (waitlist) and the docs say 'see pricing inNot published: pricing page needs sign-in (waitlist) and the docs say 'see pricing in your account'; docs state Arker charges for run time, not machine time; managed pools bill their reserved size at on-demand rates, BYOC and self-hosted pools bill VM use at 60% off on-demand rates; no free tier | — | Memory and disk persistent by default (local NVMe, remote object storage when idle)Memory and disk persistent by default (local NVMe, remote object storage when idle); policies can schedule suspend (e.g. 10 min after last start) and delete; inbound requests or runs wake a suspended VM and runs carry on where they were | VM | May 2026 | Core API primitives fork, run and sync; Python, TypeScript SDKs, CLI and REST; fork as a first-class primitiveCore API primitives fork, run and sync; Python, TypeScript SDKs, CLI and REST; fork as a first-class primitive; bare-metal fleet on AWS/GCP/Hetzner; regions incl. AWS us-west-2 and us-east-1; BYOC pools in your AWS account; self-hosted pools via open-source BLAST (Stanford); Linux, Windows and Mac support, nested virtualization, virtual GPUs; desktops with noVNC for computer-use agents; programmable network policies with credential injection; 99.9% SLA claim; pools in private access for partnered background-agent companies |
| Core | Core | Sandboxes billed on RAM held, per minute: active $0.04572/GiB-hr (forks freeSandboxes billed on RAM held, per minute: active $0.04572/GiB-hr (forks free; baking bills as sandbox time), parked ~$0.000139/GiB-hr; sandbox snapshots $0.10/GiB-month, cold $0.05/GiB-month. Sizes: sandbox-sm 1 vCPU/2 GiB (bills 2 GiB, ~$0.0914/h), sandbox-md 4 vCPU/8 GiB (bills 16 GiB, ~$0.7315/h), sandbox-lg 8 vCPU/32 GiB, sandbox-xl 16 vCPU/64 GiB. No 2 vCPU/4 GiB size. Nodes: cpu-2 (2 reserved cores, 8 GiB) $0.144/hr, cpu-8 $0.576/hr, h100-1 $4.032/hr, h100-8 $32.256/hr, b300-1 $9.792/hr; 20% weekend discount; prepaid credit, new accounts start at $0 | Parked ~$0.000139/GiB-hr on memory handed back (processes keep running)Parked ~$0.000139/GiB-hr on memory handed back (processes keep running); snapshots $0.10/GiB-month, cold $0.05/GiB-month; $0 once deleted | A sandbox runs until you delete it and bills for the RAM it holdsA sandbox runs until you delete it and bills for the RAM it holds; after 1.5 s without a command it parks itself, handing back unused memory while processes, files and sockets stay; the next command wakes it within ~10 s (tens of ms typical); snapshots can be taken live and forked (memory and processes included), and live until deleted or an optional expiry; command deadline 60 s default, 10 min max | VM | Jul 2026 | Agent-first GPU PaaS for ML research (nodes, batch jobs, sweeps, RL rollouts) with microVM sandboxes beside the GPUsAgent-first GPU PaaS for ML research (nodes, batch jobs, sweeps, RL rollouts) with microVM sandboxes beside the GPUs; bake an environment once and fork up to 256 per call (~50 ms per fork, 1,000 at once); live snapshot and branch; egress on or none; public HTTPS preview URLs for sandbox ports (24 h expiry, optional bearer token); images sbx-base, sbx-min, sbx-task or your own linux/amd64 container image; up to 4,096 sandboxes per workspace, 8 GiB RAM per sandbox raisable to 64 GiB; driven by MCP server (OAuth), HTTP API with OpenAPI, and autor CLI (formerly gman); ComputeSDK provider package @computesdk/givemeanode |
| Core | Core | $0.2168 | — | Billed from microVM boot to session end, idle included | VM | — | Code Interpreter, Browser and Runtime; interactive shells for Runtime |
| Core | Core | $0.2522 | $0 compute while suspended (up to 8 h); snapshot storage $0.08/GB-month | Suspend for up to 8 hours with memory and disk kept and no compute chargeSuspend for up to 8 hours with memory and disk kept and no compute charge; snapshots billed | VM | Jun 2026 | 2 GB of memory per vCPU baseline, can scale to 4x baseline up to 8 GB / 4 vCPU; 8-hour total runtime per MicroVM |
| Core | Core | No additional charge: you pay for the parent EC2 instancesNo additional charge: you pay for the parent EC2 instances, plus the EKS Auto Mode or ECS Managed Instances fee if used; no per-sandbox rate, so no 2 vCPU/4 GiB price | Same as the parent EC2 instance, which bills while it runs | Lives as long as its EKS pod, ECS task or VM; the VM is destroyed when the pod exitsLives as long as its EKS pod, ECS task or VM; the VM is destroyed when the pod exits; stopping, rebooting or terminating the parent instance stops its sandboxes; no suspend or pause stated | VM | Nov 2026 | Hardware isolation from the parent instance and other sandboxesHardware isolation from the parent instance and other sandboxes; cryptographic attestation via the Nitro Secure Module (confidential mode) or a direct mode without it; EKS via runtimeClassName isolated-sandbox or confidential-sandbox; ECS Managed Instances; direct use on EC2 with Amazon Linux and microVM; up to 5 sandboxes per parent vCPU, at most 896 per instance; each sandbox at most as many vCPUs as the parent, at least 256 MiB; no GPU passthrough; most Nitro-based instances, Graviton, Intel and AMD; all commercial regions |
| Core | Core | — | — | Per the SDK's types: one-shot jobs with a TTL, ephemeral sandboxesPer the SDK's types: one-shot jobs with a TTL, ephemeral sandboxes, and persistent machines in on_demand or always_on mode that can be paused; size profiles carry an idle-pause setting | — | Oct 2026 | Go SDK (Apache-2.0, standard library only, 0 stars; v0.5.0 is the first public release and the API is pre-1.0) forGo SDK (Apache-2.0, standard library only, 0 stars; v0.5.0 is the first public release and the API is pre-1.0) for projects, jobs, sandboxes, persistent machines, templates, services, execs, files, terminals, egress and inbound rules, events and webhooks; ships an in-memory test server and a contract suite; API base URL cloud.azaleva.net |
| Core | Core | $0.216 | Sandboxes: $0 CPU/memory while stopped (snapshot and volume storage rates notSandboxes: $0 CPU/memory while stopped (snapshot and volume storage rates not on the pricing page); dynamic sessions have no paused state | Sandboxes: stop or auto-suspend after a configurable idle timeoutSandboxes: stop or auto-suspend after a configurable idle timeout, snapshot in memory mode (memory + disk) or disk mode, sub-second resume; no CPU/memory fees while stopped; optional auto-delete N days after stop. Dynamic sessions: ephemeral, destroyed after the pool's cooldown period with no activity | VM | Jun 2026 | Sandboxes: sub-second start from cached images, snapshots for suspend/resume and cloning, OCI imagesSandboxes: sub-second start from cached images, snapshots for suspend/resume and cloning, OCI images, volumes (Azure Blob, data disk), egress policies with credential injection, VNet integration, ingress ports with optional Entra ID auth, Python/TypeScript/.NET SDKs and REST, tiers XS 0.25 core/0.5 GiB to 3XL 16 cores/32 GiB/320 GiB, thousands of sandboxes per sandbox group; dynamic sessions: prewarmed session pools, built-in Python/Node.js/Shell interpreter or custom container; Foundry hosted agents run Agent Framework/LangGraph agents on a managed runtime |
| Core | Core | Credits: 1 credit = 60 s x 1 CPU x 1 GiB, RAM paired 1:1 with CPU (0.5-4 cores)Credits: 1 credit = 60 s x 1 CPU x 1 GiB, RAM paired 1:1 with CPU (0.5-4 cores); Free $0 with 1,000 credits/month (1 CPU/1 GiB, 60 s max); Pro $97/mo with 10,000 credits then $1.00 per 1K credits (up to 4 CPU/4 GiB, 1 h max); Enterprise self-hosted, custom; storage 10 GB included then $0.10/GB-month. No 2 vCPU/4 GiB configuration; at the Pro overage rate 2 CPU/2 GiB = 120 credits/hr ($0.12/hr) and 4 CPU/4 GiB = 240 credits/hr ($0.24/hr) | $0 compute (nothing runs between calls)$0 compute (nothing runs between calls); thread state counts as storage, $0.10/GB-month after 10 GB | No idle state: every execute call gets a fresh sandbox that is torn down afterwardsNo idle state: every execute call gets a fresh sandbox that is torn down afterwards; optional thread_id saves /home/baponi diffs to cloud storage between calls; max execution 60 s (Free) or 1 h (Pro) | Process | 2026 | Sub-20 ms isolation setup claimed; MCP and REST API, Python SDK; Python, Node.js, BashSub-20 ms isolation setup claimed; MCP and REST API, Python SDK; Python, Node.js, Bash; BYOB storage mounts (S3, GCS, Azure) with path scoping; custom OCI images; outbound network blocked by default, 200 Mbps per core; credential-free connectors; audit logs on every tier; Enterprise deploys into your Kubernetes/VPC, air-gapped; written in Rust; founded 2026 in Austin, TX |
| Core | Core | $0.2154 | $0 compute on standby; snapshots $0.20/GB-month; volumes $0.16/GB-month | Auto-suspend to standby after about 15 s without requests (homepage)Auto-suspend to standby after about 15 s without requests (homepage), snapshotting memory, files and running processes; resumes in ~25 ms on the next request, even months later; standby billed for storage only; default expiration 7 days after creation, configurable as max age (TTL_MAX_AGE) and/or idle time (TTL_IDLE); expiry deletes the sandbox and its files; volumes persist separately | VM | Nov 2026 | Launched 2026-11-05 next to Baseten's model inference and training; boot a fresh microVM and connect in under 1 sLaunched 2026-11-05 next to Baseten's model inference and training; boot a fresh microVM and connect in under 1 s; resume from standby ~25 ms; in-memory root filesystem; default size 4 GiB / 2 vCPU, memory sets vCPUs (account max 4,096 MiB by default); up to 10 concurrent sandboxes per org by default; Agent Drive shared files and durable volumes; egress controls with credential injection into outbound requests; public or private HTTPS previews; custom images; logs and metrics in the dashboard; SDKs, CLI, Management API, sandbox runtime API, MCP; region us-pdx-1 shown in docs; fork from snapshot coming soon |
| Core | Core | Builder $20/mo (10,000 commands/mo); Pro $50/mo (100,000 commands/mo)Builder $20/mo (10,000 commands/mo); Pro $50/mo (100,000 commands/mo); Lifetime $200 once (100,000 commands/mo); free key on sign-up; checkout through Gumroad; custom limits on request | $0 (nothing persists) | Stateless: every request gets a fresh environment and files are wiped after execution | — | Jan 2026 | One POST /execute endpoint with an x-api-key header returns stdout; 5 ms boot claim; outbound internet (curl) allowedOne POST /execute endpoint with an x-api-key header returns stdout; 5 ms boot claim; outbound internet (curl) allowed; free key needs a mainstream email address (Gmail, Outlook, Yahoo); the API host in the site's own code example (bash-api.com) did not resolve on 2026-11-14 |
| Core | Core | $0.14 | $0 compute when not running; storage not billed yet (planned $0.0001/GB-hour) | Sandbox has a default 600 s lifetime (keep_warm_seconds; -1 keeps it until terminatedSandbox has a default 600 s lifetime (keep_warm_seconds; -1 keeps it until terminated; update_ttl() extends it); billed only while running; filesystem and memory snapshots, new sandboxes can fork from a snapshot; separate persistent VMs stop with no compute charge | Container | Nov 2023 | Sandboxes boot in seconds; filesystem and memory snapshots, fork from snapshot; GPUs can be attachedSandboxes boot in seconds; filesystem and memory snapshots, fork from snapshot; GPUs can be attached; Python and TypeScript SDKs; separate persistent Linux VMs for agents (microVMs); BYOC and self-hosted options; open-source engine beta9 (AGPL-3.0, 1,852 GitHub stars); SOC 2 Type II; at the Shared tier 2 vCPU/4 GiB would be $0.044/hr |
| Core | Core | Sandboxes: no published price (beta waitlist)Sandboxes: no published price (beta waitlist). CI runners per minute at 2 vCPU: Ubuntu x64 $0.004 (2 vCPU / 8 GB), Ubuntu ARM $0.0025, Windows x64 $0.008; 3,000 free min/mo; Docker layer caching and sticky disks $0.50/GB-month; static IPs $100/IP-month. Its own [code]smith agent's sandbox compute is $0.002 per vCPU-minute | — | CI runner microVMs are ephemeral and destroyed after each jobCI runner microVMs are ephemeral and destroyed after each job; the Sandboxes SDK (preview) has pause and resume; no idle policy published | VM | Oct 2026 | Sandboxes for coding agents: full VMs that run Docker and mirror CI, environment snapshots to boot from, x86 and ARMSandboxes for coding agents: full VMs that run Docker and mirror CI, environment snapshots to boot from, x86 and ARM, Linux, Windows and Mac; TypeScript SDK @useblacksmith/sandbox (preview): exec, long-lived processes, output streaming, files, exposed ports, pause/resume; main product is GitHub Actions runners on bare-metal gaming CPUs (microVMs boot in under 3 s); Testbox CLI (early beta); SDK repo MIT, 0 stars |
| Core | Core | $0.1656 | $0 compute; snapshots $0.20/GB-month | Standby: suspended time is not billed; snapshot storage billed | VM | — | Up to 256 GB RAM per sandbox on custom plans; Zero Data Retention support |
| Core | Core | Paid per request in USDC through x402 (Base, Solana or Arc)Paid per request in USDC through x402 (Base, Solana or Arc), no account needed: create from $0.011 (CPU, up to 300 s, 1 vCPU, 1 GiB); longer or larger CPU sandboxes $0.25/hour plus a $0.001 request fee; exec, status and terminate $0.002 each in the API reference ($0.001 on the service page); GPU burst create T4 $0.051 to H100 $0.401; GPU hourly T4 $4.00, L4 $4.50, A10G $5.00, A100 $7.00, H100 $9.00; CPU sandboxes max 1 vCPU/1 GiB, GPU sandboxes up to 8 vCPU/32 GiB; no 2 vCPU/4 GiB CPU size | N/a (no pause; the full requested lifetime is paid at create) | Fixed lifetime chosen at create, 10 s to 24 h (default 300 s)Fixed lifetime chosen at create, 10 s to 24 h (default 300 s), paid upfront with no refund on early terminate; each exec runs at most 60 s; no pause or suspend | — | — | Resale of Modal sandboxes to agents that pay with x402Resale of Modal sandboxes to agents that pay with x402; managed Python 3.11 image only, no custom images or setup commands; outbound network only to PyPI, Hugging Face and GitHub; 5 running sandboxes per payer (2 of them GPU); Python and TypeScript SDKs handle payment; BlockRun, Inc., backed by HF0 and NVIDIA Inception |
| Core | Core | No public price for the sandbox (early access or demo)No public price for the sandbox (early access or demo); 'BlueRock for Builders' AI workspace is free for 30 days, then $10 a month | — | — | Container | — | Agent security platform from BlueRock Security Inc. (San Mateo): a governed runtime for coding agents used from ClaudeAgent security platform from BlueRock Security Inc. (San Mateo): a governed runtime for coding agents used from Claude Code, Cursor, GitHub Copilot, Codex and VS Code, whose execution a connector relays into an isolated remote devcontainer (BlueDev Container) reached with short-lived OIDC-signed SSH certificates; inline policy on model calls, MCP tools, shell, files, network and data; audit trail and per-agent cost caps; sandbox fleets with per-team templates; deployed as EC2 (CLI, Terraform, CloudFormation), GCP or Azure images in the customer's cloud; also an MCP Trust Registry scanner, observability and guardrails |
| Core | Core | $0.018 | $0 (stopped; disk kept at least 14 days at no charge)$0 (stopped; disk kept at least 14 days at no charge); saved snapshots past plan allowance $0.056/day | No pause; stop snapshots the disk and ends billingNo pause; stop snapshots the disk and ends billing; machine held 10 min unbilled so a quick resume usually keeps memory and processes; after that disk only, kept at least 14 days; auto-stop default 1 h, up to 30 days or off on paid plans; disk saved every minute while running | VM | Jun 2026 | Ubuntu VMs with root, SSH/SCP, Docker, systemd, /dev/kvm; 60fps virtual desktop and computer use; fork from snapshotUbuntu VMs with root, SSH/SCP, Docker, systemd, /dev/kvm; 60fps virtual desktop and computer use; fork from snapshot; public IPv6 and outbound IPv4; preinstalled Docker, VS Code, Chrome, GitHub CLI, Rust, Node.js, Bun; seven agent harnesses via boat prompt (Claude Code, Codex, pi, OpenCode, Prime Agent, Kimi Code, Mistral Vibe); machines in Germany, Finland, France only; no GPU; 100-2,000 concurrent by plan; Python and TypeScript SDKs, REST API, CLI; SOC 2 in progress |
| Core | Core | Billed per active second: Small 1 vCPU / 4 GB / 20 GB $0.0219/hBilled per active second: Small 1 vCPU / 4 GB / 20 GB $0.0219/h; Medium 2 vCPU / 8 GB / 40 GB $0.0281/h; Large 4 vCPU / 16 GB / 50 GB $0.0472/h; no 2 vCPU / 4 GB size. Paused storage free for 3 days, then $0.05/GB-month. Free $0 with a $1 one-time trial (one Small at a time); Pro $10, $20 or $50 a month of included usage (unlocks Medium and Large, up to 25 concurrent); Enterprise custom. The execution API and eval model tokens draw on the same allowance | Paused: $0 compute; storage free for 3 days, then $0.05/GB-month | Pause releases CPU and RAM and keeps files (compute billing stops)Pause releases CPU and RAM and keeps files (compute billing stops); resume starts the same workspace; per-sandbox idle_timeout and max_lifetime options (platform defaults not stated); fork copies /workspace but not running processes | — | Aug 2026 | Python and TypeScript SDKs, CLI (macOS, Linux, WSL) and a desktop app; exec and interactive terminalPython and TypeScript SDKs, CLI (macOS, Linux, WSL) and a desktop app; exec and interactive terminal; public URLs for ports; environments with language runtimes or coding-agent images; RL environment pools scored by your own verifier; custom RL sandboxes with sub-5 ms steps on reserved capacity; Judge0-style execution API (Python, Node, Go, C, C++) with NumPy, SciPy and pandas preloaded; evals with hosted models (DeepSeek, Kimi, GLM, MiMo, LongCat); agent tracer; GPU sandboxes and CI runners coming soon; checkout shows INR amounts |
| Core | Core | vCPU and memory metered only while running; no published ratesPricing page | $0 compute (storage not stated) | Stop when idle, snapshot, resume later without rebuildingStop when idle, snapshot, resume later without rebuilding; stopped and suspended sandboxes incur no compute charge | — | — | Persistent sandboxes with terminals, files, services and browsers; snapshots and workspace transferPersistent sandboxes with terminals, files, services and browsers; snapshots and workspace transfer; managed capacity or private host pools; web app, API, CLI and MCP server |
| Core | Core | Fixed sizes, billed per second: Small 0.5 vCPU / 1 GiB about $0.05/hFixed sizes, billed per second: Small 0.5 vCPU / 1 GiB about $0.05/h; Large 1.5 vCPU / 3 GiB about $0.15/h; XL 4 vCPU / 8 GiB about $0.40/h and XXL 4-8 vCPU / 16 GiB about $0.80/h (Starter only); Free $5 credit for 14 days; pay as you go $5-$500 credit; Starter $15.99/month founding price with $40 of compute; no 2 vCPU / 4 GiB size | $0 while stopped | A VM stops after 30 minutes idle and keeps its workspace filesA VM stops after 30 minutes idle and keeps its workspace files; billed per second while running, idle time included, with a one-minute minimum; Starter keeps computers on but bills at most 100 hours a month | VM | Sep 2026 | Full Ubuntu VM with 30 GiB disk, ready in about 30 s; bring your own agent over MCP (Claude Code, Codex, CursorFull Ubuntu VM with 30 GiB disk, ready in about 30 s; bring your own agent over MCP (Claude Code, Codex, Cursor, Claude) or use BoxCompute's own agent; desktop you can watch and take over (computer use); live preview URLs; GitHub integration; Vault for secrets kept out of the agent's view; HTTP API (v2) and integrations for LangGraph, CrewAI, Mastra, Pi, Hermes; network closed by default, private ranges and SMTP port 25 blocked; servers in Cyberjaya, Malaysia |
| Core | Core | $0.15 | Hibernated: disk only, $0.000096/GiB-h on data written (EUR 0.008/h at 100 GiBHibernated: disk only, $0.000096/GiB-h on data written (EUR 0.008/h at 100 GiB written); standby: RAM + disk | Keeps running when you disconnect; after 4 hours with no network traffic it hibernatesKeeps running when you disconnect; after 4 hours with no network traffic it hibernates (memory written to disk, disk-only billing) and wakes on the next request with processes intact (156 ms); standby (pause) keeps memory warm, resumes in under 1 ms; runtime unlimited; optional delete-after timer | VM | May 2026 | Persistent Linux VMs, default 2 vCPU/8 GiB/100 GB disk; ready in <10 msPersistent Linux VMs, default 2 vCPU/8 GiB/100 GB disk; ready in <10 ms; fork a running machine with memory and disk in <200 ms; snapshots and in-place checkpoints; SSH; HTTPS domain per machine, custom domains, TCP/UDP port forwards; desktop view and computer-use API; Docker preinstalled; integrations (GitHub, Linear, Slack) and MCP server for Claude Code, Codex, opencode; egress allowlist and secrets swapped in on the way out; regions Frankfurt and New York; BYOC/self-host on AWS, GCP or any KVM host; Python and TypeScript SDKs; CLI and macOS menu bar |
| Core | Core | — | — | The CLI has sandbox timeouts and pause and resume commands; limits and billing not stated | VM | Nov 2026 | E2B-compatible API: the stock E2B SDKs and CLI work with E2B_DOMAIN=boxfab.ioE2B-compatible API: the stock E2B SDKs and CLI work with E2B_DOMAIN=boxfab.io; JavaScript (npm boxfab@alpha) and Python SDKs derived from E2B's (MIT); alpha CLI with templates built from Dockerfiles; code interpreter (Python, JavaScript, TypeScript, R, Java, Bash); desktop sandbox template for computer-use agents (Apache-2.0); the README says 'BoxFab is a product of ContextFab, Inc.'; GitHub org created 2026-11-01, 0 stars, pushed 2026-11-14 |
| Core | Core | Per second: browser session $0.045/h; a session with a shell is billed by its machine atPer second: browser session $0.045/h; a session with a shell is billed by its machine at $0.045 per vCPU-hour + $0.013 per GiB-hour, and there is one machine size, 1 vCPU / 3 GiB (about $0.084/h), so no 2 vCPU / 4 GiB price; stopped $0; no per-session fee. Plans: Free $0 ($5 credit a month, 3 concurrent, 15-minute sessions), Hobby $19/mo ($15 included, 25 concurrent, 2 h), Startup $99/mo ($80 included, 100 concurrent, 6 h), Scale custom (250+, 24 h). Residential proxies $8/GB, datacenter $1.50/GB, CAPTCHA solving $3 per 1,000 | $0 while stopped (kept 7-90 days by plan) | Timeout defaults to 300 s, at most 15 min (Free), 2 h (Hobby)Timeout defaults to 300 s, at most 15 min (Free), 2 h (Hobby), 6 h (Startup) or 24 h (Scale); optional idleTimeout; a browser-only session stops when its last CDP client disconnects unless keepAlive is set; stop saves the whole browser and /workspace (not processes or memory) and frees the machine; stopped sessions are kept 7, 30, 30 or 90 days by plan, then deleted | gVisor | Sep 2026 | Real Chrome over CDP (Playwright, Puppeteer) with live view and human takeover, proxies, CAPTCHA handlingReal Chrome over CDP (Playwright, Puppeteer) with live view and human takeover, proxies, CAPTCHA handling, browser profiles and recordings; bash shell with Python, Node and sudo, alone or next to the browser on the same machine; /workspace disk 10 GiB (5 GiB browser-only) plus about 4 GiB for the system; no GPU; AI agent runs plus fetch, crawl and search APIs; credentials vault; hosted MCP server, Node and Python SDKs, CLI; 70 examples; warm pool targeting under 1 s to a running session |
| Core | Core | $0.1584 | Stopped: disk only, $0.00018/GiB-h (Medium 20 GiB is $2.59 per 30 days) | Cloud: idle-stop with wake-on-access (autostop seconds, 0 disables)Cloud: idle-stop with wake-on-access (autostop seconds, 0 disables); a stopped box keeps its disk and pays for it until deleted. Local runtime: you manage the lifecycle | VM | Dec 2025 | Open source Apache-2.0 (boxlite-ai/boxlite, 2,646 stars), embeddable libraryOpen source Apache-2.0 (boxlite-ai/boxlite, 2,646 stars), embeddable library; macOS Apple Silicon, Linux x86_64/ARM64, WSL2; Cloud per-box max 4 vCPU/32 GiB/120 GiB; preview URLs; egress allowlist; secret injection; managed volumes not metered; Python and TypeScript SDKs, CLI, REST API (104 endpoints); Cloud images base, python, node |
| Core | Core | Managed: $0.30 per active hour for one size, 4 vCPU / 8 GiB RAM / 40 GiB workspaceManaged: $0.30 per active hour for one size, 4 vCPU / 8 GiB RAM / 40 GiB workspace, billed per second; $0 while asleep; prepaid credit, minimum top-up $10, no subscription. Brezel runtime is free to self-host. No 2 vCPU / 4 GiB size | $0 compute; workspace kept until expiry (no storage fee stated) | Sleep (standby) stops compute and processes and keeps the /workspace until the sandboxSleep (standby) stops compute and processes and keeps the /workspace until the sandbox expires; wake is automatic or manual; per-sandbox TTL (expires_after_seconds) and standby-after timer; 'resident compute' for always-on services | VM | Sep 2026 | HTTP API and dashboard; Python SDK brezel-sdk, npm @infercrane/brezel, CLIHTTP API and dashboard; Python SDK brezel-sdk, npm @infercrane/brezel, CLI; commands, files, temporary preview URLs (port leases), browser connect; model keys kept outside the sandbox, sandboxes offline unless a policy allows; self-run StarSling Node CPU benchmark 29.90 runs/s; self-host needs an Ubuntu 24.04 x86-64 host with KVM; not yet: OCI builds, interactive PTY/SSH, desktop, multi-node, GPU passthrough; Apache-2.0, 1 GitHub star; InferCrane's main business is inference operations |
| Core | Core | $0.062 | $0 compute while paused; persistent volume kept (rate ?)$0 compute while paused; persistent volume kept (rate ?); snapshots $0.25/GB-month | Manual pause scales the container to zero and stops CPU/memory billingManual pause scales the container to zero and stops CPU/memory billing; only a persistent volume survives a pause (no RAM or processes); optional auto-destroy after 30m to 18h or when the main process exits (oneShot); platform max-lifetime ceiling | gVisor | May 2026 | Ready in about 2-3 s (create blocks until ready); templates for Node, Python, BunReady in about 2-3 s (create blocks until ready); templates for Node, Python, Bun, Deno plus coding-agent templates (e.g. Factory Droid); snapshots of the filesystem; persistent volumes 10-50 GB; preview URLs on brimble.run; egress can be blocked; TypeScript, Python and Go SDKs and REST API; region iad-1 shown; part of a wider app-hosting platform (deploys, databases, object storage, CI runners) |
| Core | Core | Billed per second in euro: pico 0.25 shared vCPU/128 MB €0.003/h, micro €0.004/hBilled per second in euro: pico 0.25 shared vCPU/128 MB €0.003/h, micro €0.004/h, nano 1/512 MB €0.006/h, small 1/2 GB €0.016/h, medium 2 shared vCPU/4 GB €0.03/h, large 4/8 GB €0.058/h, xlarge 8/16 GB €0.12/h (Pro); disk and egress not billed today; plans Hobby free, Pro €20/mo (€20 usage included), Team €100/mo (€100 included); €10 free credit; no USD list price (dashboard can display USD at the day's rate) | $0 compute while sleeping or stopped$0 compute while sleeping or stopped; disk not billed today (within the plan's disk allowance) | Sleeps after its sleep window without use (default 1 hSleeps after its sleep window without use (default 1 h; 5 min to 7 days, or never on Pro and above); a sleeping briq keeps its disk, is not billed for compute, and the next call or URL visit wakes it from a memory snapshot in a second or two; nothing is deleted because of time | VM | Oct 2026 | Startup about 300 ms, first command about 100 ms; sandboxes with shell, desktop, browser, nested VM and AndroidStartup about 300 ms, first command about 100 ms; sandboxes with shell, desktop, browser, nested VM and Android; services (Postgres, Redis, DuckDB, any OCI image) on the same private network; fork; MCP server (in Claude's connector directory), CLI (@briq-cloud/cli), REST API; regions eu-west (Roubaix, France) default and us-east (Beauharnois, Canada); open-source microVM runtime briqyard (AGPL-3.0, 4 stars) |
| Core | Core | $0.01 per hour of compute, billed in tokens (a live sandbox costs 10 tokens per hour$0.01 per hour of compute, billed in tokens (a live sandbox costs 10 tokens per hour; top-ups $10 per 10,000 tokens). Free $0/mo: 10,000 tokens (up to 1,000 h), personal and non-commercial use only. Pro $20/mo: 20,000 tokens (up to 2,000 h), commercial use, then $0.01 per extra hour. Enterprise custom with self-hosting and SSO. Compute runs on the end user's device, so no CPU/RAM size is sold | $0 (code and files stay on the user's device; no storage charge) | Runs on the device that opened it, while the tab is openRuns on the device that opened it, while the tab is open; each instance's filesystem (OPFS or IndexedDB) is persistent between visits or ephemeral, per instance | WebAssembly | Feb 2026 | In-browser sandboxes from Leaning Technologies (the team behind WebVM and CheerpX)In-browser sandboxes from Leaning Technologies (the team behind WebVM and CheerpX); Node.js and Rust runtimes, Python in preview, Go and Ruby planned; npm, git and bash inside the pod; Portals expose a port in the pod on a public URL; outbound network requests are proxied through BrowserPod infrastructure; unlimited concurrent sandboxes and no storage charge; npm package @leaningtech/browserpod; up to 10M tokens for startups and 100k tokens/month for open-source projects |
| Core | Core | $0.0864 | $0 compute; disk $0.00011/GB-hour (~$0.08/GB-month) | Optional inactivity timeout (none by default, no lifetime cap) stops the sandboxOptional inactivity timeout (none by default, no lifetime cap) stops the sandbox; stop ends running processes (memory isn't kept), the disk persists; a stopped sandbox wakes on traffic to an endpoint or on a command; Buddy deletes nothing on its own schedule; snapshots survive deletion | VM | Apr 2017 | 70 ms start claim; 1 vCPU / 2 GB up to 12 vCPU / 24 GB; 50 GB persistent disk; Ubuntu 22.04 or 24.0470 ms start claim; 1 vCPU / 2 GB up to 12 vCPU / 24 GB; 50 GB persistent disk; Ubuntu 22.04 or 24.04; root, Docker inside, SSH and VS Code Remote-SSH; preinstalled node, docker, go, python3, git, Claude Code, Codex and Cursor CLI; MCP server, bdy CLI, TypeScript SDK, REST API, YAML; HTTP/TCP/TLS endpoints and custom domains; regions include us-1 and eu-1; part of the Buddy CI/CD platform |
| Core | Core | Pay per run in USDC on Solana (x402), no account or API keyPay per run in USDC on Solana (x402), no account or API key; fixed sizes priced live, not published as a table (homepage example: small-10m for 0.05 USDC); a paid order that starts no sandbox within 30 minutes is refunded; in-sandbox x402 API spending capped at 0.50 USDC a call and 5 USDC per sandbox; GPU sizes (H100, A100, L40S, L4) coming soon | N/a (no pause; deleted at the deadline) | Fixed-length runs: each size sets CPU, memory, disk and minutesFixed-length runs: each size sets CPU, memory, disk and minutes; the sandbox and everything in it are deleted at the deadline; more time can be bought with bugnet_extend; stopping early deletes it and unused time is not refunded; no pause | — | Nov 2026 | MCP server (mcp.bugnet.sh) for Claude Code, claude.ai, Cursor, VS Code, Codex CLI and DevinMCP server (mcp.bugnet.sh) for Claude Code, claude.ai, Cursor, VS Code, Codex CLI and Devin, plus a Claude Code plugin and a local bridge that can pay from the pay.sh wallet or a wallet file; the agent assesses a repo and quotes a price, and the user approves every paid call; output stripped of escape codes and labeled as data (last 8 KiB per stream); file transfer by one-use curl links; SSH, web previews, secrets and GPU on request; Linux only; open-source (Apache-2.0) MCP server, bridge and plugin, GitHub bugnet-sh/bugnet 3 stars |
| Core | Core | CPU Functions and CPU Sandboxes: memory $0.00001667/GB-s ($0.06/GB-hour)CPU Functions and CPU Sandboxes: memory $0.00001667/GB-s ($0.06/GB-hour), requests $0.20/1M, egress $0.09/GB, and no vCPU rate (a 4 GB CPU Sandbox would cost $0.2400/hour, GB treated as GiB, but the docs don't say how many vCPUs a CPU Sandbox gets); GPU Functions and GPU Sandboxes: vCPU $0.0000126/vCPU-s, memory $0.00000137/GB-s, GPU $0.00013067/GPU-s ($0.4704/GPU-hour), egress $0.09/GB; model storage $0.39/GB-month; Free $0 with $5/mo of CPU usage included; Pro $20/mo or $192/yr, including $20 of usage and 20 GB of model storage; Enterprise contact sales | N/a (no paused state; a sandbox is destroyed at its timeout) | No pause or suspend: a sandbox runs until you call delete() or its timeout endsNo pause or suspend: a sandbox runs until you call delete() or its timeout ends, then it is destroyed; default timeout 1 minute; sandboxes can run up to 24 hours | VM | May 2025 | Serverless platform for agents: CPU and GPU Functions (top-level apps) spawn nested CPU and GPU Sandboxes for untrustedServerless platform for agents: CPU and GPU Functions (top-level apps) spawn nested CPU and GPU Sandboxes for untrusted actions such as AI-generated code or RL reward functions; CPU Sandboxes run Python, Go, Node.js, Deno or Bash; GPU Sandboxes run Python with automatic model mounting and concurrent requests on one GPU (example GPU type T4G), up to 50 vCPU and 75 GB memory; sandboxes can also be created from code running outside Buildfunctions; Runtime Controls (budgets, loop and circuit breakers, policy gates; beta) work standalone; TypeScript and Python SDKs (Apache-2.0, 2 GitHub stars each); Enterprise adds on-premise options |
| Core | Core | Plans set concurrency: Free (5 sandboxes/day, 1 concurrent, 30-min timeoutPlans set concurrency: Free (5 sandboxes/day, 1 concurrent, 30-min timeout, 7-day history); Pro $29/month (unlimited sandboxes, 5 concurrent, 4-hour timeout); Team $99/seat/month (15 concurrent, shared). Compute is metered per minute on top, by size: 1 vCPU/256 MB $0.001/min, 2 vCPU/512 MB $0.002, 2 vCPU/1 GB $0.003 ($0.18/h), 4 vCPU/2 GB $0.005, 4 vCPU/4 GB $0.008 ($0.48/h), 8 vCPU/8 GB $0.015; no 2 vCPU/4 GiB size (rates from the site's cost estimator, 'list rates as of 20 September 2026'); LLM tokens metered per model at its egress proxy; disk and egress not billed | $0 compute while asleep (caged sleep stops billing); disk not billed | A sandbox sleeps after an idle timeout you set (timeout key in .caged.yamlA sandbox sleeps after an idle timeout you set (timeout key in .caged.yaml, clamped to the plan's maximum); caged sleep pauses it and stops billing, caged wake resumes it; plan timeouts Free 30 min, Pro 4 h; snapshots restore into another sandbox; destroy deletes the VM and its rootfs overlay, while the session record is kept | VM | Jun 2026 | Runs Claude Code, Aider, OpenAI Codex CLI, Grok CLI, Cline, Continue CLI and Goose, or any MCP agentRuns Claude Code, Aider, OpenAI Codex CLI, Grok CLI, Cline, Continue CLI and Goose, or any MCP agent; records eight event types per session for replay; behaviour trust score; Personas, Shifts and Teams with autonomy tiers; pipelines with approval stages and A2A hand-off; network modes full, allowlist or none; templates minimal, node-22, node-20, python-312, python-311 and desktop; limits 8 vCPU, 8192 MB and 50 GB disk per sandbox; REST API, MCP over WebSocket, CLI, Python and TypeScript SDKs (MIT); company Bytangle Ltd |
| Core | Core | No pricing published; cantelop.com/pricing returns 404Pricing page | — | A sandbox lasts for one activation; keepAliveSeconds keeps it warm between messagesA sandbox lasts for one activation; keepAliveSeconds keeps it warm between messages (memory and temp state kept); when keep-alive expires the sandbox is released and its memory, processes and root-filesystem changes are lost; Session ID and the per-user /workspace files (NFS) persist, and the next message starts a fresh sandbox; session.stop() releases the sandbox at once | VM | Sep 2026 | Each user gets a durable Workspace with ephemeral SessionsEach user gets a durable Workspace with ephemeral Sessions; runs unmodified Claude Code, Codex, OpenCode, Pi or your own harness; users can bring their own Claude/Codex subscriptions; TypeScript SDK 1.0 (typed Apps); CPU and memory grow and shrink with the workload; egress limited to configured ports, private ranges and cloud metadata blocked; first-time Workspace activation 0.99 s median (their benchmark); open source: platform AGPL-3.0, SDK MIT (github.com/stepandel/cantelop, 0 stars, repo created 2026-10-21) |
| Core | Core | $0.10 per minute ($6.00/hour) per Mac session, billed from started_at to ended_at$0.10 per minute ($6.00/hour) per Mac session, billed from started_at to ended_at; prepaid $5 packs (500 credits = 50 minutes); 5 free credits on sign-up; optional $5 auto top-up; machine size (cores, memory) not stated | $0 after stop (no paused state) | Sessions are time-boxed (max_minutes default 60, max 480)Sessions are time-boxed (max_minutes default 60, max 480); stop_session ends billing and the Mac session is gone; queue wait is free; unused reserved minutes are refunded as credits; no standing Mac is kept | VM | Aug 2026 | Disposable Mac sessions for agents and scripts: Xcode, Swift, iOS simulators, codesign, notarizationDisposable Mac sessions for agents and scripts: Xcode, Swift, iOS simulators, codesign, notarization, AppleScript and Shortcuts; you can sign into your own Apple ID in a session; MCP with OAuth 2.1 (create_session, exec, upload_file, tail_logs, download_file, stop_session), REST API and TypeScript SDK (@caprun/sdk, not on npm yet); founder's post says it is built on Capgo's builder tech |
CastariSunsetting: Castari Cloud is winding down and the repo is being open-sourced Core | Core | Free during the public beta; usage-based pricing for compute was planned for GAFree during the public beta; usage-based pricing for compute was planned for GA; Castari Cloud is now sunsetting | N/a (no idle state; sandboxes reset after each invocation) | Fresh sandbox per invocation: a warm sandbox runs the agent and its state is resetFresh sandbox per invocation: a warm sandbox runs the agent and its state is reset afterwards, so nothing persists between invocations; a redeploy destroys the old sandbox; 120 s timeout per invocation by default | Container | Jan 2026 | Deploys Claude Agent SDK projects with one command (cast deploy) from the CLI or a GitHub importDeploys Claude Agent SDK projects with one command (cast deploy) from the CLI or a GitHub import; default limits 2 shared CPU cores, 2 GB memory, 10 GB ephemeral disk; MCP gateway, model router, tool traces and logs; secrets, managed files and storage buckets mounted into sandboxes; TypeScript SDK; CLI MIT (GitHub castari/cli, 7 stars); the team now builds madethis.com |
CathedralStandard sandboxes closed to new orders since 2026-10-02 Core | Core | $0.035 | Same as running (a paused sandbox is billed until deleted) | Runs until deleted, its stop time (lifetime_seconds up to 24 hRuns until deleted, its stop time (lifetime_seconds up to 24 h, extendable) or its spend cap; console stop times 30 min, 1, 4 or 23 h; pause/resume freezes it but stays billed | VM | Jul 2026 | Signed receipt for every command (verifiable offline); attestation per host, not per room; E2B SDK compatibleSigned receipt for every command (verifiable offline); attestation per host, not per room; E2B SDK compatible; up to 1,000 sandboxes at once, 600 starts/min; 10 GiB disk; no SSH into sealed sandboxes; desktop/computer use and browser use 'coming soon'; sealed snapshots behind a flag; repo describes it as Bittensor SN94 |
| Core | Core | $0.03 | $0 compute when stopped; disk kept 30 days (storage price not shown) | Stop keeps the root disk (stopped disks kept 30 days), processes endStop keeps the root disk (stopped disks kept 30 days), processes end; pause/resume keeps processes; snapshots of memory, disk and processes; command timeout 30 s default, 300 s max | VM | Feb 2026 | Cold start under 2 s (cloud), about 500 ms (open-source package); terminal over WebSocket (celesto computer ssh)Cold start under 2 s (cloud), about 500 ms (open-source package); terminal over WebSocket (celesto computer ssh); live desktop/VNC on browser templates; Chromium over CDP; forking; up to 4 public HTTPS ports; Docker inside; network off option; Windows 11 and macOS preview in the OSS package; Python and TypeScript SDKs, CLI, MCP server; GitHub Actions runners; open source Apache-2.0 (github.com/CelestoAI/celesto, 1.4k stars) |
| Core | Core | $0.2454 | $0 compute while suspended (suspend releases CPU, memory and GPUs)$0 compute while suspended (suspend releases CPU, memory and GPUs); snapshot storage price not stated; volumes $0.09/GiB-month | No automatic idle suspend stated; a sandbox runs until terminated or until its optionalNo automatic idle suspend stated; a sandbox runs until terminated or until its optional lifetime expires; suspend() saves memory and filesystem (gVisor checkpoint) and releases CPU, memory and GPUs, and restore() resumes it under the same ID; snapshot() forks new sandboxes; the root filesystem is discarded at termination while volumes persist; non-preemptible workloads run up to 24 h | gVisor | Mar 2026 | Python SDK chalkcompute and chalk CLI over gRPC; default image Python 3.14 with Git and SSHPython SDK chalkcompute and chalk CLI over gRPC; default image Python 3.14 with Git and SSH; cold start under 2 s for common base images (docs); egress network policy by host, CIDR and port, with secret injection; managed SSH and tunnels for ssh, scp and remote IDEs; durable versioned volumes; GPUs (H100, H200, B200, B300, L4, L40S, A100, A10), with CUDA state in snapshots; MCP Gateway; coding-agent helpers (Claude Code, OpenCode); runs in your own cloud (self-hosted) or on Chalk Serverless; part of Chalk's data and AI platform; SDK repo chalk-ai/chalk-sandbox-sdk (no license file, 1 star) |
| Core | Core | $0.18 | $0 while asleep; disk is discarded unless saved (Container snapshots free$0 while asleep; disk is discarded unless saved (Container snapshots free during beta) | An instance stops after an inactivity timeout you set (setInactivityTimeout)An instance stops after an inactivity timeout you set (setInactivityTimeout); no fixed maximum runtime; charges stop when it sleeps; disk is ephemeral, so the next start uses a fresh disk unless you saved a Container snapshot (beta) or mounted R2; cold starts often 1-3 s | VM | Jun 2025 | Started and controlled from a Worker's Durable Object; any image you provideStarted and controlled from a Worker's Durable Object; any image you provide; internet access on or off per instance, outbound handlers; Container snapshots (beta); R2 mounts via FUSE; Dynamic Workers for JS/Python/Wasm code; default limits 4 vCPU/12 GiB/20 GB per instance and 15,000 concurrent vCPU per account (can be raised); instances placed across Cloudflare's network; Sandbox SDK on GitHub (cloudflare/sandbox-sdk, 1,165 stars); coding-agent runner guide |
cloudrouter (Manaflow) CoreAppears discontinued: cloudrouter.dev redirects to manaflow.comAppears discontinued: cloudrouter.dev redirects to manaflow.com, which no longer mentions it | Core | No pricing page; cloudrouter.dev now redirects to manaflow.com | — | Stop pauses a sandbox and resume restarts itStop pauses a sandbox and resume restarts it; sandboxes have a timeout that cloudrouter extend lengthens (length not stated); delete removes it permanently | VM | Feb 2026 | CLI (npm @manaflow-ai/cloudrouter, also cr) and agent skill for Claude Code, Codex and CursorCLI (npm @manaflow-ai/cloudrouter, also cr) and agent skill for Claude Code, Codex and Cursor; sizes small 2 vCPU/8 GB/20 GB, medium 4 vCPU/16 GB/40 GB, large 8 vCPU/32 GB/80 GB (default); GPUs T4 to B200; Chrome with CDP browser automation, VNC desktop, VS Code in the browser, Jupyter Lab, PTY, file upload and download with watch; Docker support; CLI source in manaflow-ai/manaflow (MIT); manaflow.com now lists only cmux, its terminal for coding agents |
| Core | Core | Private beta via waitlist; no sandbox price on the pricing pagePricing page | N/a (no pause; deleted at TTL) | Ephemeral: create, exec, delete; nothing persistsEphemeral: create, exec, delete; nothing persists; TTL default 60 min, maximum 480 min, then the sandbox is removed; Sandbox.using() and the Python context manager delete it automatically | VM | Jul 2026 | About 125 ms boot (site claim); flavors python-3.12, node-20, browser (Chromium, Playwright, xvfb) and full (adds jqAbout 125 ms boot (site claim); flavors python-3.12, node-20, browser (Chromium, Playwright, xvfb) and full (adds jq, ffmpeg, ImageMagick); default 512 MB and 1 vCPU, up to 8192 MB and 4 vCPU; REST API with TypeScript (@codecapsules/sandbox) and Python (codecapsules-sandbox) SDKs, MIT; streaming exec, file upload and download; tool definitions for Claude, OpenAI and LangChain; part of the Code Capsules PaaS |
CodeBox APIHosted API appears shut down (codeboxapi.com does not resolve on 2026-11-14) Core | Core | Subscription plans gave an API key; the pricing page at codeboxapi.com/pricing no longerSubscription plans gave an API key; the pricing page at codeboxapi.com/pricing no longer resolves | — | — | Container | Jul 2023 | Python package codeboxapi for running Python code in a sandbox from LLM apps: exec, package installPython package codeboxapi for running Python code in a sandbox from LLM apps: exec, package install, file upload and download, asyncio support, local mode for development without internet; used by codeinterpreter-api; GitHub shroominic/codebox-api 363 stars, MIT; last PyPI release 0.2.6 (2025-01-30) |
| Core | Core | Local VMs free; cloud per second while runningLocal VMs free; cloud per second while running, EUR only: €0.0438/vCPU-hour + €0.0141/GB-hour memory + €0.000087/GB-hour disk; presets 1 vCPU/4 GB, 2 vCPU/8 GB (€0.2091/h with 100 GB disk), 4/16, 8/32, or custom up to 16 vCPU/32 GB/200 GB; 2 vCPU/4 GiB compute = €0.1440/h (GB treated as GiB); no USD price published; prepaid balance with €20 top-ups; Custom/self-hosted by contact | €0 while stopped (pricing: nothing while it's stopped; disks kept) | Stop powers the VM off and keeps its disks; a stopped hosted sandbox is not billedStop powers the VM off and keeps its disks; a stopped hosted sandbox is not billed; optional TTL deletes a hosted sandbox (cw extend adds hours; cw protect blocks deletion); if the balance runs out, running sandboxes stop after 24 h and storage is kept | VM | Oct 2026 | Same cw CLI for local and hosted VMs; encrypted private mesh between sandboxesSame cw CLI for local and hosted VMs; encrypted private mesh between sandboxes, with cw msg for agent-to-agent messages; Claude Code and Codex integrated; Docker inside each VM; credential proxy; preview URLs; SSH and VS Code; TypeScript, Python and Go SDKs; hosted in Amsterdam and Frankfurt on bare metal, US coming; self-host on Kubernetes; written in Rust; control plane and CLI AGPL-3.0, SDKs MIT; GitHub codewiresh/codewire 1 star |
| Core | Core | $0.128 | $0 compute when suspended; snapshot storage $0.10/GB-month | Suspend stores a snapshot and stops the CPU/RAM meter; resume keeps memorySuspend stores a snapshot and stops the CPU/RAM meter; resume keeps memory; maximum session 1 hour (Demo), 24 hours (Pro), unlimited (Enterprise) | VM | — | Live fork of a running microVM with process memory in single-digit milliseconds (~2.4 ms p50), up to 1,000 per groupLive fork of a running microVM with process memory in single-digit milliseconds (~2.4 ms p50), up to 1,000 per group; built for RL post-training rollouts and coding agents; REST API, Python SDK, hosted MCP; per-fork egress; deterministic replay; Pro up to 100 concurrent (site) or no concurrency limit (docs); Enterprise dedicated nodes, SSO, BYOC |
| Core | Core | Current preview fleet (public pricing API): one fixed size, 2 vCPU / 2000 MiB / 20 GiBCurrent preview fleet (public pricing API): one fixed size, 2 vCPU / 2000 MiB / 20 GiB, at $0.03/vCPU-hour + $0.01/GiB-hour = $0.0795/hour, paid from purchased credits; site's proposed launch pricing: Usage $0.030/vCPU-hour + $0.010/GiB-hour with no platform fee; Build $20/month with $25 credits; Committed $250/month minimum at 15% off; Dedicated custom; public billing and checkout not available yet. Workspace snapshots $0.021/GB-month + $6.30 per million writes + $0.504 per million reads; proposed retained storage $0.08/GiB-month and egress $0.05/GiB after 100 GiB | Snapshots $0.021/GB-month; proposed retained storage $0.08/GiB-month | Fleet VMs are ephemeral leases of 1 to 30 minutes (default 10)Fleet VMs are ephemeral leases of 1 to 30 minutes (default 10), with files and installed tools gone when the VM is deleted or expires; workspace snapshots keep files past a lease; pricing FAQ says persistent-tier sandboxes can be paused to end compute charges, but the docs say this release has no automatic pause/resume | VM | Sep 2026 | Sandbox layer shared by Condensation's own services (Codegraff agents, nanohub Git)Sandbox layer shared by Condensation's own services (Codegraff agents, nanohub Git); preview VM has Bun, Python 3.14 with uv, Zig, Rust and Git, with Codegraff, Codex, Claude and Cursor agents; npm and Python SDKs (condensationai 0.2.0) and a local MCP bridge; two active VMs per account; named HTTPS address per app preview; workspaces.condensation.ai (the input's address) now redirects to a broken path on condensation.ai |
| Core | Core | Cloud (launch pricing): $20/seat/month flat, no usage bill or overageCloud (launch pricing): $20/seat/month flat, no usage bill or overage; a seat is one box with 2 vCPU (1 guaranteed, bursts to 2) / 4 GiB RAM / 20 GB disk / 100 GB transfer per month, and the box sleeps when idle; OSS self-host: your VM's cost; teams, dedicated hardware and own-cloud by conversation | Flat $20/seat/month whether awake or asleep | Boxes are long-lived: disk, installs and caches survive until you delete the boxBoxes are long-lived: disk, installs and caches survive until you delete the box; idle boxes sleep (Cloud); failed CI boxes kept alive 1 h for debugging, extendable | Container | Dec 2025 | One Linux box per agent with real root; MCP server inside every box (shell and file ops) for Claude Code, CursorOne Linux box per agent with real root; MCP server inside every box (shell and file ops) for Claude Code, Cursor, Cline etc.; routable hostname with TLS, SSH via built-in sentinel (sshpiper + Caddy); GitHub Action containarium-run for CI with keep-on-failure; NVIDIA GPU passthrough on your own hardware; self-host on one Ubuntu VM, air-gapped bundle; open source Apache-2.0 (GitHub FootprintAI/containarium, 318 stars); by FootprintAI; site in English, Traditional Chinese, Japanese |
| Core | Core | $0.3408 | Same as running while it exists (no idle timeout)Same as running while it exists (no idle timeout); snapshot and volume storage billed separately | No idle timeout: a sandbox runs until you stop it or its lifetime expires and bills itsNo idle timeout: a sandbox runs until you stop it or its lifetime expires and bills its requested CPU and memory the whole time; filesystem snapshots restore or fork into a new sandbox, so compute can stop without losing the workspace | VM | May 2026 | Part of CoreWeave Forge; Python and TypeScript clients over a protobuf v1 APIPart of CoreWeave Forge; Python and TypeScript clients over a protobuf v1 API; burst from your CKS cluster to managed capacity; egress limits by IP, port and HTTPS hostname; single- and multi-container (Docker Compose) workloads; PVC volumes; built for RL rollouts, evals and agent tool use |
| Core | Core | $0.1187 | Paused: $0 vCPU; memory $0.01159025/GiB-h and storage $0.0001584/GB-h stillPaused: $0 vCPU; memory $0.01159025/GiB-h and storage $0.0001584/GB-h still billed | Pause is a microVM snapshot (memory, registersPause is a microVM snapshot (memory, registers, devices) to durable storage: vCPU billing stops, memory and disk still billed; auto-pause on idle off by default (60-86,400 s); no maximum session length; paused sandboxes count toward the concurrency cap | VM | Jun 2026 | ~210 ms median time to interactive; fork; private networks between sandboxes; egress allowlist~210 ms median time to interactive; fork; private networks between sandboxes; egress allowlist; disk sync to your S3, MinIO or R2; Dockerfile templates; devbox:1 image with Claude Code, Codex, OpenCode, Cursor CLI and Pi preinstalled; desktop:1 image with computer-use API; SSH; TypeScript SDK, CLI, MCP, REST; EU and US; self-host on own infra (Enterprise); no GPU; shapes up to 8 vCPU/16 GB (Enterprise); Free plan max 1 vCPU/1 GB; ISO 27001, SOC 2 Type II |
CubeShutting down (sunset notice 2026-11-06; final deletion 2026-12-20) Core | Core | $0.0164 | $2 per vCPU per month paused ($4/month for 2 vCPU) | Persistent cloud computer that runs 24/7 until pausedPersistent cloud computer that runs 24/7 until paused; pause keeps files but not memory or processes; out of credit pauses it | — | Aug 2026 | Shutting down: announced 2026-11-06, no new machines, remaining machines deleted 2026-12-20Shutting down: announced 2026-11-06, no new machines, remaining machines deleted 2026-12-20; Mac app (Apple Silicon), web app and phone; SSH from VS Code, Cursor, Zed, Ghostty; Claude Code, Codex, Gemini CLI, Copilot CLI, Cursor CLI, OpenCode, goose and Kimi Code with the user's own accounts; up to 16 GB burst RAM; by Collaborators Inc. (also makes Collab and Proton) |
| Core | Core | No charge during Beta; Databricks expects to bill later for runtime computeNo charge during Beta; Databricks expects to bill later for runtime compute, persistent storage and data transfer (rates not published) | $0 during Beta; future storage billing announced without a rate | Stop and start at any time; home directory (up to 100 GB) persistsStop and start at any time; home directory (up to 100 GB) persists, the 10 GB outside it is deleted on stop; inactivity auto-stop (duration not stated, can be disabled with --no-autostop); each start runs at most 24 h, then Databricks stops it | — | — | Fixed size 4 vCPU / 8 GB RAM; REST API, Databricks SDK for Python (recommended) and CLI; SSH for CLI-created sandboxesFixed size 4 vCPU / 8 GB RAM; REST API, Databricks SDK for Python (recommended) and CLI; SSH for CLI-created sandboxes; custom container images from Artifact Registry; sandbox gets its own Databricks credential to call workspace APIs, workspace files and Unity Catalog volumes; use cases include running an agent harness and AI-written code; requires a Databricks workspace in a supported region; beta data may be purged at end of Beta |
| Core | Core | Credits, 100 credits = $1 ("provisional")Credits, 100 credits = $1 ("provisional"). Size classes billed per second: small 1 CPU/2 GiB 0.0008 credits/s (~$0.029/h), medium 4 CPU/8 GiB 0.0016 credits/s (~$0.058/h), large 8 CPU/16 GiB 0.0032 credits/s (~$0.115/h). AWS Marketplace public offer: small $0.03/h, medium $0.06/h, large $0.12/h, any GPU $4.80/h. GPU classes on Modal or Daytona are billed by that provider. No 2 vCPU/4 GiB class; nearest are small (1 CPU/2 GiB) and medium (4 CPU/8 GiB). The /pricing page loads its rates from a signed-in API and could not be read | Same as running while the sandbox runs (credits burn even when idle)Same as running while the sandbox runs (credits burn even when idle); $0 once terminated | You reserve a run time when you launch a sandbox and that time is deducted from yourYou reserve a run time when you launch a sandbox and that time is deducted from your credits up front; ending it early refunds the unused part; a running sandbox burns credits even when idle; runtime snapshots can be saved and loaded into new sandboxes | — | — | Agent studio and notebook platform for data agents, built on JupyterAgent studio and notebook platform for data agents, built on Jupyter; connect Claude Code, Codex, Cursor or any MCP client to a notebook or sandbox (Datalayer MCP server; first code execution binds a sandbox to the session); applications get a Computer with shell, files and preview browser where Claude Code, Codex, OpenCode or pi run; user environments built from Datalayer bases with apt/conda packages or a Dockerfile step; CUDA base for GPU versions on Modal/Daytona; Agent Nodes on your own AWS account; browser (Pyodide) and local sandboxes free; open-source pieces on GitHub (jupyter-mcp-server, code-sandboxes BSD-3-Clause, agent-runtimes) |
| Core | Core | $0.1656 | Storage $0.000108/GiB-hr beyond 5 GiB freeStorage $0.000108/GiB-hr beyond 5 GiB free; compute cost while stopped not stated | Stopped/archived cost not stated on pricing page | — | — | Per-second billing; Windows option; startup credits up to $50k |
| Core | Core | Lite 0.5 vCPU/512 MB/5 GB $0.005/h; Basic 1 vCPU/1 GB/10 GB $0.01/hLite 0.5 vCPU/512 MB/5 GB $0.005/h; Basic 1 vCPU/1 GB/10 GB $0.01/h; Standard 2 vCPU/2 GB/20 GB $0.02/h; Large 4 vCPU/4 GB/40 GB $0.04/h; described as active CPU pricing (active CPU only), metered every 5 minutes from a prepaid USD balance funded through Stripe, minimum balance 1 hour of the chosen size; homepage also mentions x402 USDC payments; no 2 vCPU/4 GB size | $0 while sleeping (data kept) | Idle sandboxes sleep automatically (timeout not stated) and wake on the next requestIdle sandboxes sleep automatically (timeout not stated) and wake on the next request; sleep and wake API; when the balance runs out the sandbox auto-sleeps with no data loss; snapshots, fork and restore | Container | — | Full Linux environment with shell, filesystem and background processesFull Linux environment with shell, filesystem and background processes; images: Ubuntu 22.04 base, Python, Node, full (adds Go, Rust) or any Docker image; exec, files, git clone, processes, port exposure with preview URLs, WebSocket PTY; JS/TS (@dcdn/sandbox) and Python (dcdn-sandbox) SDKs; WireGuard mesh networking; runs on 8+ independent nodes in 4 regions, with 70% of revenue going to node operators; part of DCDN Cloud (decentralized CDN, DNS, WAF, GPU marketplace) |
| Core | Core | $0.149 | $0 compute; storage beyond allowance $0.0001/GiB-hour (about $0.073/GiB-month) | Autosleep after 5 minutes idle by default (configurable, or never)Autosleep after 5 minutes idle by default (configurable, or never); sleep stops processes and compute billing; the root filesystem persists, while processes, memory, network connections and /tmp don't (docs); storage stays billable while asleep | VM | Jul 2026 | 50 ms boot via snapshot restore (site claim); full Linux with systemd and root50 ms boot via snapshot restore (site claim); full Linux with systemd and root; GPU support and nested virtualization (site FAQ); POSIX filesystem (100% pjdfstest); disaggregated compute and storage with live migration; default machine 1 vCPU / 4096 MiB / 10 GiB; SSH access; CLI (dedalus 0.5.0) and TypeScript, Python and Go SDKs; startup credits up to $10,000; access needs the organization to be admitted to the Machines service |
| Core | Core | Per-second billing, no minimum, while creating, starting, running or stoppingPer-second billing, no minimum, while creating, starting, running or stopping; plan sizes from nano at $0.054/hour, default medium; per-plan specs and rates only in the signed-in catalog | $0 (a stopped sandbox costs nothing) | Idle timeout (default 1 h, set per sandbox, cannot be disabled) stops itIdle timeout (default 1 h, set per sandbox, cannot be disabled) stops it; stop keeps only /workspace and start boots a fresh container from the base image; auto-stop after 24 h running; stopped sandboxes deleted after 7 days; failed sandboxes deleted in ~5 min without /workspace | VM | Aug 2026 | Python SDK and HTTP API; exec, run_python, file read/write inside /workspace (100 MiB per call)Python SDK and HTTP API; exec, run_python, file read/write inside /workspace (100 MiB per call); outbound internet, no inbound connections, 200 Mbit/s caps, SMTP blocked; 25 non-stopped sandboxes per account; 30 min per-command cap; snapshots and expose_port on the roadmap |
Deno Sandbox CoreClosed to new customers (Nov 2026 update)Closed to new customers (Nov 2026 update); Deno Deploy shuts down April 2027 as the Deno team joins Cloudflare | Core | $0.164 | No paused state; volume storage $0.20/GiB-month | No suspend; default lifetime is the client session (microVM shuts down when the scriptNo suspend; default lifetime is the client session (microVM shuts down when the script closes it); duration timeouts keep it alive after disconnect, extendable, up to 30 minutes; volumes and snapshots for persistence; can be promoted to a Deno Deploy app for longer life | VM | Feb 2026 | Start under 200 ms (93 ms shown); 2 vCPU, 768 MB-4 GB memory (1.2 GB default), 10 GB ephemeral diskStart under 200 ms (93 ms shown); 2 vCPU, 768 MB-4 GB memory (1.2 GB default), 10 GB ephemeral disk; regions ams (Amsterdam) and ord (Chicago); secrets replaced by placeholders and injected only for approved hosts; outbound allowlist through a proxy; expose HTTP, SSH, VS Code editor; deploy a sandbox to Deno Deploy; SDKs for JavaScript/TypeScript (npm, JSR) and Python; volumes and snapshots; 5 concurrent sandboxes per org by default during pre-release |
| Core | Core | $0.36 | No paused state; snapshot storage price ? | No pause: a stopped sandbox can't be restartedNo pause: a stopped sandbox can't be restarted; timeoutMinutes defaults to 120, maximum lifetime 24 h; snapshot() saves the home directory and /workspace (not processes or memory) to restore into a new sandbox | VM | Oct 2026 | Default 2 vCPU / 4096 MB / 100 GB disk, 2 to 64 vCPUs, x86-64 or ARM64; Docker inside; Node.js SDK @depot/sandboxDefault 2 vCPU / 4096 MB / 100 GB disk, 2 to 64 vCPUs, x86-64 or ARM64; Docker inside; Node.js SDK @depot/sandbox; Depot CI secrets, Tailscale tailnet join, Depot Cache; custom images; durable logs; the same sandboxes run Depot CI and Managed Agents (Claude Code, Claude Managed Agents, OpenAI, Devin, Amp) |
| Core | Core | $0.078 | MicroVMs paused: $0 compute; storage $0.05/GiB-month on data storedMicroVMs paused: $0 compute; storage $0.05/GiB-month on data stored; checkpoints billed until deleted. Harness Runtime paused: $0 compute; checkpoint storage $0.05/GiB-month | MicroVMs auto-pause after an idle timeout (default 5 min, up to 4 h) keeping memoryMicroVMs auto-pause after an idle timeout (default 5 min, up to 4 h) keeping memory, files and processes; next HTTP request, command or console session resumes it; manual pause/resume; checkpoints capture memory and disk. Harness Runtime: pausing stops compute charges and keeps files and conversation history; workspace saved on removal or checkpoint | VM | Oct 2026 | MicroVMs run a container image (linux/amd64) in their own kernel; checkpoint to start new MicroVMs already warmed upMicroVMs run a container image (linux/amd64) in their own kernel; checkpoint to start new MicroVMs already warmed up; manage via Control Panel, doctl and API; VPC support (NAT gateway for internet); region MKC1. Harness Runtime: managed service for agent harnesses with built-in tools such as Chromium; Claude Code named as an example harness; harness: none option; five shapes 1/1 to 16/32 |
| Core | Core | Self-hosting free (AGPL-3.0, every feature)Self-hosting free (AGPL-3.0, every feature); Dock Cloud coming soon (waitlist), to be metered per vCPU-hour, GB-hour of memory, GB-month of disk and snapshots, and GB of egress, with no rates published; stopped docks pay only for disk | Disk only (stopped docks use no CPU or memory; disk rate not published) | Stop releases compute and keeps the disk; resume brings back the same id and filesStop releases compute and keeps the disk; resume brings back the same id and files; branch copies a prepared dock (stopping a running source first for a consistent copy); delta snapshots in S3-compatible storage | VM | Sep 2026 | Five verbs: new, exec/ssh, stop, resume, branch; dock port gives an HTTPS URL per portFive verbs: new, exec/ssh, stop, resume, branch; dock port gives an HTTPS URL per port; dock push and pull move a dock between laptop and control plane; hosts dial out to the control plane, with no inbound ports; connectors for GitHub, Figma, Linear, Slack and Sentry, plus an LLM gateway that keeps model keys out of the dock; REST API, CLI and TypeScript, Python and Go SDKs (installed from GitHub); Ubuntu with root and Docker; GitHub usedock/dock 2 stars, AGPL-3.0; built in India |
| Core | Core | $0.14 | $0 (a paused sandbox costs nothing; volumes free) | Local: sandbox persists until removed. Cloud: expires after 1 hour by defaultLocal: sandbox persists until removed. Cloud: expires after 1 hour by default, extendable to at most 24 h from creation; on timeout resumable sandboxes are stopped and the rest deleted; stop keeps memory and filesystem and resume keeps the ID; templates can capture disk, or memory + disk + microVM checkpoint; compute not billed while stopped | VM | Jan 2026 | Sbx CLI for macOS, Windows, Linux; agents: Claude Code, Codex, Copilot, Cursor, Devin, Docker Agent, Droid, KiroSbx CLI for macOS, Windows, Linux; agents: Claude Code, Codex, Copilot, Cursor, Devin, Docker Agent, Droid, Kiro, OpenCode, shell; sbx move carries a sandbox between local and cloud; credentials injected by a host-side proxy, never inside the VM; network, filesystem and MCP policies; SSH integration for VS Code and Cursor; Sandboxes API and SDK; Kits for reusable environments; local GPU passthrough |
| Core | Core | $0.1656 | $0 compute when paused (no storage charge stated) | Pause keeps filesystem and memory (running processes survive)Pause keeps filesystem and memory (running processes survive); paused sandboxes kept indefinitely with no TTL; optional auto-pause on inactivity; max continuous runtime without a pause 1 h on Hobby, 24 h on Pro, reset by pause/resume | VM | Mar 2023 | Open source Apache-2.0 (e2b-dev/E2B 14,863 stars; e2b-dev/runtime 1,843 stars); pause/resume and fork with memoryOpen source Apache-2.0 (e2b-dev/E2B 14,863 stars; e2b-dev/runtime 1,843 stars); pause/resume and fork with memory; snapshots; desktop sandbox for computer use; BYOC in AWS, GCP or Azure (Enterprise); Python and JS/TS SDKs; up to 8 vCPU/8 GiB self-serve; SOC 2, HIPAA; '2B+ sandboxes started' claim |
| Core | Core | $0.1944 | $0 (site: the sandbox stops when the run does$0 (site: the sandbox stops when the run does; 'nothing idles, nothing is stored, nothing keeps billing') | The sandbox stops when a turn ends and resumes for the next turn with the same filesystemThe sandbox stops when a turn ends and resumes for the next turn with the same filesystem and checkout; if the box is gone a replacement boots and the session snapshot is restored; unclaimed pre-warmed sandboxes are deleted after about 10 minutes | — | Aug 2026 | Agents as an API: POST a prompt and Epho boots Claude Code, CodexAgents as an API: POST a prompt and Epho boots Claude Code, Codex, OpenCode or Copilot in a cloud sandbox with your repos, files and MCP servers attached, streaming events over SSE; sizes 2 or 4 vCPU and 2, 4 or 8 GiB memory, 10 GiB disk; structured output schemas; webhooks; artifacts and per-repo git diffs returned; pre-warm endpoint; no raw command-exec endpoint; by Bruin Data Limited (Bruin) |
| Core | Core | $0.132 | No user pause described; a balance-exhaustion pause to snapshot is not billed | Ephemeral (default): deleted on kill, after an idle timeout (5 min defaultEphemeral (default): deleted on kill, after an idle timeout (5 min default, up to 1 h) or when its TTL ends (up to 24 h). Persistent: keeps disk and state for up to 30 days and is billed per started hour while running; if the balance runs out it is paused to a snapshot (not billed) and kept 14 days | VM | Jun 2026 | Ready in about a second (site claim); Python 3.12, Node.js 22, bash, git and curlReady in about a second (site claim); Python 3.12, Node.js 22, bash, git and curl; outbound internet, no SSH or inbound ports; Python (eqvps) and TypeScript (@eqvps/sdk) SDKs and hosted MCP server (mcp.eqvps.com), MCP repo Poiuyhje/eqvps-mcp (MIT, 0 stars, 75 tools); email-only no-KYC signup, agents can register over MCP; daily and monthly spending caps; same balance as EQVPS's VPS plans (servers in Germany and Finland) |
| Core | Core | Developer pod, flat monthly: S 4 vCPU / 8 GB / 80 GB $24/monthDeveloper pod, flat monthly: S 4 vCPU / 8 GB / 80 GB $24/month; M 8 vCPU / 16 GB / 160 GB $42/month; L 16 vCPU / 32 GB / 320 GB $79/month; no 2 vCPU / 4 GiB size; bring your own model plan or key. Managed OpenClaw pod $29/month including up to $10 of model usage (bigger plans from $59/month) | Same as running (flat monthly plan, always on) | Always on; no suspend; the disk is backed up daily (about 7 days of rotation)Always on; no suspend; the disk is backed up daily (about 7 days of rotation); after cancellation the pod runs to the end of the paid period, then the machine and backups are deleted after 7 days | — | Oct 2026 | Ubuntu 24.04 pod with you as administrator, with Docker, git, Node, Python, Chrome, ffmpeg and tmuxUbuntu 24.04 pod with you as administrator, with Docker, git, Node, Python, Chrome, ffmpeg and tmux; Claude Code, Codex, OpenCode, Pi or Hermes installed, or any agent; reached only over your Tailscale network (SSH through Tailscale); ready in 10 to 15 minutes; Everpod API and hosted MCP server to order pods; managed OpenClaw over Telegram, Discord, Slack, WhatsApp, Signal; made by DEH Technologies LLC (Wyoming) |
| Core | Core | $0.105 | — | — | — | — | Pools shared across many VMs; 100 GiB filesystem included |
| Core | Core | $0.1645 | $0 compute when warm or cold; cold storage $0.000027/GB-hour (~$0.02/GB-month) | About 30 s after activity stops the Sprite goes warm: VM suspended with memory frozenAbout 30 s after activity stops the Sprite goes warm: VM suspended with memory frozen, no compute billing, resumes in 100-500 ms with processes intact. Later it goes cold: VM stopped, memory dropped, disk kept, wakes in 1-2 s with processes started fresh. A request to its URL wakes it; checkpoints are taken automatically | VM | Jan 2026 | Fixed 8 vCPUs and 100 GB disk per Sprite, memory managed and auto-scaled by the platformFixed 8 vCPUs and 100 GB disk per Sprite, memory managed and auto-scaled by the platform; create in about 1 s (launch post); live copy-on-write checkpoints and restore; HTTPS URL per Sprite that wakes it; Connectors to GitHub, Slack, OpenRouter or any HTTP API keep credentials outside; services restart after a cold wake; CLI, REST, JavaScript, Go, Elixir, Python SDKs; MCP; Hero plan runs 100 Sprites plus 100 warm at once |
| Core | Core | No pricing page left: forevervm.com redirects to modal.comPricing page | — | README: idle machines are swapped from memory to disk automatically and swapped back whenREADME: idle machines are swapped from memory to disk automatically and swapped back when needed, so REPLs run 'forever' | — | Feb 2025 | API for stateful Python REPL machines; CLI (npx forevervm) and JS SDK @forevervm/sdk; per-machine memory limitsAPI for stateful Python REPL machines; CLI (npx forevervm) and JS SDK @forevervm/sdk; per-machine memory limits; MIT, 230 GitHub stars, last push 2025-04-17; maker Jamsocket joined Modal on 2025-07-10, whose post says Jamsocket products continue while core functionality moves to Modal |
| Core | Core | $0.1322 | — | Paused VM cost not stated | — | — | Saved VMs; concurrency limits by plan |
| Core | Core | $0.0137 | $0 running time; stopped disk kept free for 14 days$0 running time; stopped disk kept free for 14 days; saved snapshots beyond the plan's allowance $0.056/day | Runs until stopped or its auto-stop timer ends (1 h after each start by defaultRuns until stopped or its auto-stop timer ends (1 h after each start by default, up to 30 days, or none; counts from start, not activity); stop keeps disk 14 days; start within 10 min of a stop resumes running programs (warm resume), later starts boot fresh | VM | Nov 2026 | About 20 s from API call to SSH; Ubuntu 24.04/26.04, Debian 13 or Omarchy with Docker, ChromeAbout 20 s from API call to SSH; Ubuntu 24.04/26.04, Debian 13 or Omarchy with Docker, Chrome, a browser desktop and ten coding agents (Claude, Codex, Gemini, OpenCode, pi, Amp, Cursor and more); automatic hourly snapshots, named snapshots, fork; HTTPS URL per port and public IPv6; froog CLI, HTTP API; also sells Pushbox GitHub Actions runners ($0.0006/vCPU-min) and Tokenmail inboxes ($2/mo); no GPU, no BYOC |
| Core | Core | Sandboxes need the Pro plan ($125/month per organization, flat, unlimited usersSandboxes need the Pro plan ($125/month per organization, flat, unlimited users, 14-day trial) or Enterprise (custom, on-premise); early access and opt-in; each sandbox 1 CPU and 1 GiB memory; at most 10 concurrent per organization; no per-hour sandbox price | $0 (destroyed after 30 minutes idle) | Ephemeral Ubuntu 24.04 sandboxes are destroyed automatically after 30 minutes ofEphemeral Ubuntu 24.04 sandboxes are destroyed automatically after 30 minutes of inactivity; exec, read and write calls reset the timer; can be destroyed from the dashboard or CLI | VM | — | Part of Gatana, an MCP gateway between agents and tools (SSO, RBAC, audit logs, SIEM, external vaults)Part of Gatana, an MCP gateway between agents and tools (SSO, RBAC, audit logs, SIEM, external vaults); sandbox image has Node.js 24, Python 3, uv, git, build tools, database clients and the Gatana CLI, signed in automatically by OIDC token exchange; exec (NDJSON stream), read-file and write-file API; web terminal and direct SSH with a 5-minute JWT; all sessions recorded in the audit log; LangChain Deep Agents backend (gatana-langchain); Gatana Cloud in the EU or on-premise on Kubernetes; Accendo Consulting GmbH, Zurich |
| Core | Core | The API's usage endpoint returns a rate card in CPU core-seconds, memory GB-secondsThe API's usage endpoint returns a rate card in CPU core-seconds, memory GB-seconds, disk GB-seconds and credits, but no rates are published in the SDK repo; needs a Genspark plan that includes SB-Box, plus credits | $0 compute while frozen (SDK README); storage cost of frozen memory and disk ? | Every sandbox has an idle timer and an optional deadlineEvery sandbox has an idle timer and an optional deadline; at timeout it is frozen (default) or killed (lifecycle.onTimeout); freeze keeps memory and disk with no compute billed; resume takes about 1 s with processes still running; frozen memory images can expire (then reset); rollback to saved versions and clone are supported | — | Nov 2026 | Official SDKs for Python (genspark-sbbox), TypeScript (@genspark/sb-box), Go and Rust, all MIT, same surfaceOfficial SDKs for Python (genspark-sbbox), TypeScript (@genspark/sb-box), Go and Rust, all MIT, same surface; REST API at genspark.ai/api/sb-box/v1 with gsk- API keys; machine ready in about a second; templates built from an image or Dockerfile (default 1 CPU, 2048 MB, 20 GB disk); exec with SSE streaming, a stateful code interpreter (Python by default), background commands, PTY and WebSocket terminal, file streaming; public port URLs; network allow and deny lists and a public-traffic token; sharing grants for templates and sandboxes; per-sandbox metrics; regions endpoint (default westus2); per-plan limits on concurrency, CPU, memory and disk; repo has 0 GitHub stars |
| Core | Core | $0.0452 | Stopping keeps the volume; whether a stopped instance still pays the monthlyStopping keeps the volume; whether a stopped instance still pays the monthly plan is not stated | Always on for a flat monthly plan; you can stop an instance and restart it later with itsAlways on for a flat monthly plan; you can stop an instance and restart it later with its persistent volume preserved; no auto-sleep stated | VM | Jun 2026 | Agent catalog: Claude Code, Aider, OpenClaw, Hermes, browser-use; HTTPS subdomain with automatic TLS per instanceAgent catalog: Claude Code, Aider, OpenClaw, Hermes, browser-use; HTTPS subdomain with automatic TLS per instance; embedded terminal; REST API (create, list, stop, start, exec, log streaming) and gentity CLI (MIT, github.com/gentityapp/cli); volumes encrypted at rest, in the iad (US East) region; bandwidth 50-400 GB and log retention 7-30 days by plan; upgrades only (volumes cannot shrink); operated by ROOTTEAM, Inc. |
| Core | Core | $0.17 | $0 during Early Access (paused storage and snapshots)$0 during Early Access (paused storage and snapshots); future storage rate not published | Default lifetime 24 h, renewing automatically unless timeout_seconds pins it (minimum 60Default lifetime 24 h, renewing automatically unless timeout_seconds pins it (minimum 60 s); pause stops run time and keeps files on disk but not memory or running processes (resume is a cold start); paused sandboxes kept up to 7 days, then deleted with their disk; snapshots of running sandboxes can seed many new sandboxes | Container | Jun 2026 | Agent hosting plus Agent Marketplace with Verified listings; Agentbox SDK for Python (gmi-agentbox-sdk) and GoAgent hosting plus Agent Marketplace with Verified listings; Agentbox SDK for Python (gmi-agentbox-sdk) and Go; commands, file transfer, interactive shell (Beta); snapshots; 200+ models through one OpenAI-compatible key on the same network; customer case claims up to 50,000 concurrent sandboxes; no Docker inside a sandbox; Agentbox v2 (2026-10-22) added persistent project workspaces and pause |
| Core | Core | Golem Cloud free during Developer Preview (resource limits apply, not for productionGolem Cloud free during Developer Preview (resource limits apply, not for production, no SLA); paid usage-based Cloud planned, rates not published; planned meters: compute in Golem Compute Units (1 GCU = 1 million fuel units of WebAssembly execution), active memory GiB-seconds, active durable and ephemeral filesystem GiB-months; self-hosting free under the repo's license; Cloud On-Prem offer planned | $0 during Developer Preview; planned meters stop accruing while idle or$0 during Developer Preview; planned meters stop accruing while idle or suspended | Durable execution: a waiting agent is suspended using no CPU or memoryDurable execution: a waiting agent is suspended using no CPU or memory, for minutes or months, and after a crash, restart or rescale resumes mid-function with memory, call stack, files and environment restored; resource meters stop while idle or suspended; Developer Preview has no data-retention guarantee | WebAssembly | Aug 2023 | Durable runtime for agents and their tools; VM starts in as little as 2 msDurable runtime for agents and their tools; VM starts in as little as 2 ms, code can be tens of kilobytes (site claims); fresh VM per tool call that can't read agent memory; tool calls run exactly once, recorded results reused on recovery; bundled Shell, Git, Node.js, npm and TypeScript tools; HTTP and MCP retries with idempotency keys; serves agent artifacts (files, pages, simple apps); Golem CLI and SDKs; open source repo golemcloud/golem (Business Source License 1.1, 1,519 GitHub stars); Golem 1.6 current |
| Core | Core | $0.206 | Agent Platform: ? (storage $0.000410959/GiB-hourAgent Platform: ? (storage $0.000410959/GiB-hour; idle-sandbox compute rule not stated); GKE: suspended Pod compute $0, volumes and snapshots billed as storage; Gemini API: $0 during preview | Agent Platform: sandbox TTL 7 days (custom container image) to 14 days (code execution)Agent Platform: sandbox TTL 7 days (custom container image) to 14 days (code execution), state can be resumed days later, snapshot APIs exist. GKE: setting operatingMode Suspended deletes the Pod but keeps the Sandbox and its volumes; memory and filesystem changes lost unless a Pod snapshot (gVisor only) is taken; no idle detection. Cloud Run: sandboxes ephemeral, deleted on completion; detached sandboxes reusable. Gemini API: auto-snapshot and stop after 15 min idle, resumable by ID for 7 days, then deleted | gVisor | Aug 2025 | GKE Agent Sandbox is open source under Kubernetes SIG Apps (kubernetes-sigs/agent-sandbox, Apache-2.0GKE Agent Sandbox is open source under Kubernetes SIG Apps (kubernetes-sigs/agent-sandbox, Apache-2.0, 4,610 GitHub stars) with Sandbox, SandboxTemplate, SandboxClaim and SandboxWarmPool CRDs; Agent Platform Code Execution has no internet or VPC access, template sandboxes have internet off by default, Computer Use gives a browser environment; Cloud Run sandboxes started with the sandbox command inside a job/instance (Preview); Gemini API environments run Antigravity and custom agents, start in about 5 s, fixed CPU and memory |
| Core | Core | $0.1256 | $0 compute while paused; storage $0.0001/GiB-h for paused runtimes$0 compute while paused; storage $0.0001/GiB-h for paused runtimes, private templates and snapshots | Runtime runs until kill() or its timeout (set per runtime in seconds)Runtime runs until kill() or its timeout (set per runtime in seconds), which terminates it and removes its local root disk; pause (on request) frees CPU and memory and keeps processes, open files, REPL state and disk, and resume brings the same runtime back; snapshots of running or paused runtimes; block-storage volumes outlive the runtime | VM | Mar 2026 | Agent Sandbox, Agent Browser (headed or headless) and Agent Computer (Linux desktopAgent Sandbox, Agent Browser (headed or headless) and Agent Computer (Linux desktop; Windows and macOS planned) on one API; ready in ~25 ms from a template or snapshot; fork from a live sandbox or snapshot; code (Python, Node), commands, PTY, files, Git; mount S3 buckets and Git repos at boot; SSH gateway and web terminal; public port URLs; network policies by host and port; MCP gateway keeps credentials outside the sandbox; managed on AWS us-east-1 and ap-south-1, Azure and Google Cloud planned; bring-your-own-cloud and on-prem (Dell, Cisco, HP AI Factory) on the roadmap; Python and TypeScript SDKs, CLI, REST; Tier 0 quota 10 vCPU, 20 GiB RAM, 50 GiB disk |
| Core | Core | No separate sandbox price; code runs on one-off dynosNo separate sandbox price; code runs on one-off dynos, which Heroku bills per second of use at dyno rates (Common Runtime: Basic 0.5 GB $7/mo, Standard-1X 0.5 GB $25/mo, Standard-2X 1 GB $50/mo, Performance-M 2.5 GB $250/mo; Fir-generation Private Spaces only: Compute dyno-2c-4gb $150/mo); the built-in code_exec tool's dyno size is not stated; Managed Inference and Agents model usage billed separately | $0 (dyno ends after each run) | Ephemeral: a one-off dyno is spun up on demand for each run and terminates after useEphemeral: a one-off dyno is spun up on demand for each run and terminates after use; runtime_params.max_calls caps runs per agent loop | Container | May 2025 | Built-in tools code_exec_python, code_exec_node, code_exec_ruby and code_exec_go in the Heroku Managed Inference andBuilt-in tools code_exec_python, code_exec_node, code_exec_ruby and code_exec_go in the Heroku Managed Inference and Agents API (/v1/agents/heroku), which install packages and retry on errors; the same as open-source MCP servers (heroku/mcp-code-exec-python, -node, -ruby, -go; CC0) for Claude Desktop, Cursor or Agentforce; related dyno_run_command and postgres_run_query tools run in one-off dynos of a target app |
| Core | Core | $0.0137 | Same as running (flat monthly, no inactivity pause) | Paid Hivra Cloud computers stay on and are never paused for inactivityPaid Hivra Cloud computers stay on and are never paused for inactivity; a computer stops when its plan stops being paid (48 h grace after a failed card payment) and keeps its files for a limited time | — | Jun 2026 | Gives each agent its own computer: Hermes Agent, Claude Code, Codex, OpenClaw, Agent Zero, or a plain Ubuntu desktopGives each agent its own computer: Hermes Agent, Claude Code, Codex, OpenClaw, Agent Zero, or a plain Ubuntu desktop; open from a browser or phone, desktop or terminal; sign in with Claude/ChatGPT accounts, API keys or Hivra credits; open source Apache-2.0, 5 GitHub stars; formerly HermesOS; separate $HIVRA crypto token (not required); site counts 2,155 agents and computers launched |
| Core | Core | $0.0671 | Same as running (flat monthly plan) | Suspends when idle and wakes in about 1 s (0.73-1.37 s measured) with filesSuspends when idle and wakes in about 1 s (0.73-1.37 s measured) with files, processes and memory kept; after days idle it may cold-start instead (about a minute, files kept); a hard restart keeps files but ends running sessions; short-lived volume snapshots only, no backups | VM | Aug 2026 | Six agent CLIs preinstalled (Claude Code, Codex, Gemini, Cursor, OpenCode, Pi) plus 60+ tools (Node 24, Python 3.11Six agent CLIs preinstalled (Claude Code, Codex, Gemini, Cursor, OpenCode, Pi) plus 60+ tools (Node 24, Python 3.11, Playwright with Chromium, Postgres 17 client, Jupyter, ffmpeg); root on your own VM; terminal in a browser tab from any device, including phones; HTTPS on your own hostname behind an auth gateway; keys stay on the box; one workstation per customer, unlimited projects; hosts the open-source HolyCode container (MIT, 84 stars; sibling HolyClaude 2,661 stars) |
| Core | Core | $0.1656 | Pricing page says 'no idle charges'; whether paused sandboxes pay the storagePricing page says 'no idle charges'; whether paused sandboxes pay the storage rate is not stated | Pause preserves full state including memory and resume continues where it left offPause preserves full state including memory and resume continues where it left off; stop keeps the filesystem but loses runtime state; optional timeout (timeout_seconds) destroys the sandbox and its data when it expires; site says no execution time limit, sandboxes can run for days or weeks | VM | Nov 2025 | ~100 ms cold start from prebuilt snapshots (site claim); Python, JavaScript, Bash and Go execution~100 ms cold start from prebuilt snapshots (site claim); Python, JavaScript, Bash and Go execution; Python and JavaScript SDKs, CLI, REST API, MCP server; desktop automation with VNC, screenshots, mouse and keyboard; preview URLs for any port; WebSocket terminal; custom templates from Docker images; background processes; SDK repo hopx-ai/hopx (MIT, 292 GitHub stars) |
| Core | Core | CNY only, billed in CUs: code interpreter 1 CU per vCPU-hour + 0.1 CU per GiB-hour (sameCNY only, billed in CUs: code interpreter 1 CU per vCPU-hour + 0.1 CU per GiB-hour (same factors for the runtime and the browser tool); CU price ¥1.1108 up to 55,000 CU a month, ¥1.0000 to 280,000, ¥0.9000 above, so 2 vCPU / 4 GiB = 2.4 CU/h = ¥2.6659/h at the first tier; free Trial edition with a one-off 20 CU; team and enterprise editions (example prices ¥119/month Team Standard with 80 CU/month, ¥1,190/month Team Advanced with 780 CU) are not yet open for purchase; no USD price list found | ¥0 after the session stops or times out (billing ends at stop-session) | Code interpreter sessions end automatically after session_timeout seconds withoutCode interpreter sessions end automatically after session_timeout seconds without activity (60-86,400, default 900 = 15 min); billing runs from the start-session call to the stop-session call | VM | — | Huawei Cloud's enterprise agent platform (智果 AgentArts): Managed AgentsHuawei Cloud's enterprise agent platform (智果 AgentArts): Managed Agents, Agent Runtime for your own agent images (LangChain, LangGraph, Google ADK), memory, gateway, observability and evaluation; code interpreter runs code, file operations and system commands, created in the console or through the runtime SDK, with a public or VPC inbound gateway and public or private outbound network settings; browser tool with sessions, human takeover and profiles; runtime default spec 2 vCPU / 8 GB (others by ticket), up to 1,000 runtimes per account; sub-second cold start claimed; Chinese-language site only (no international product page) |
| Core | Core | Billed as HF Jobs hardware, per minute: CPU Basic 2 vCPU / 16 GB $0.01/hBilled as HF Jobs hardware, per minute: CPU Basic 2 vCPU / 16 GB $0.01/h, CPU Upgrade 8 vCPU / 32 GB $0.03/h, CPU XL 16 vCPU / 124 GB $1.00/h, CPU Performance 32 vCPU / 256 GB $1.90/h; GPUs from T4 small $0.40/h to 8x H200 $40/h; exposed ports +$0.01/h per Job; needs a positive credit balance. No 2 vCPU / 4 GiB size | $0 (billing only while a Job is starting or running) | idle_timeout (default 10 min, configurable or off) shuts a sandbox down when no API callidle_timeout (default 10 min, configurable or off) shuts a sandbox down when no API call arrives and no process runs; fixed 24 h maximum lifetime; no pause or snapshot | VM | Jul 2026 | Python Sandbox API and hf sandbox CLI in huggingface_hub (Apache-2.0, 4,053 GitHub stars)Python Sandbox API and hf sandbox CLI in huggingface_hub (Apache-2.0, 4,053 GitHub stars); any Docker image with /bin/sh; about 6 s cold start and about 100 ms per command; GPU flavors on dedicated sandboxes; background processes and a port proxy; SandboxPool packs many CPU sandboxes for RL rollouts |
| Core | Core | $0.15 | Same as running while the sandbox runsSame as running while the sandbox runs; $0 compute once stopped (snapshot storage price not stated) | No pause; a sandbox runs until stopped or until timeoutMinutes (total lifetime from VMNo pause; a sandbox runs until stopped or until timeoutMinutes (total lifetime from VM start; default is the team's Session Timeout setting), and idle time is billed; memory snapshots capture memory, processes and filesystem of a running CPU sandbox for new sandboxes; GPU sandbox disks are ephemeral; macOS (beta) keeps a persistent disk but not memory | VM | Dec 2024 | Two products: cloud Chrome sessions (Playwright, Puppeteer, Selenium, stealth, CAPTCHA solving, profiles, live viewTwo products: cloud Chrome sessions (Playwright, Puppeteer, Selenium, stealth, CAPTCHA solving, profiles, live view, recordings, Web Bot Auth) and sandboxes; CPU sandbox start under 50 ms claim, browser start under 500 ms; set vCPU, memory and disk per sandbox; base and custom images; memory snapshots; exposed ports and runtime URLs; outbound network controls; secrets and vaults; GPU sandboxes (L4 to B300, on request); macOS desktop sandboxes (beta, on request); regions such as us-west; Node and Python SDKs and hx CLI; Web API, HyperAgent and computer-use agents; claims 100M+ sessions run |
| Core | Core | No public pricing; contact engineering (contact@i2cw.com)No public pricing; contact engineering (contact@i2cw.com); deployment on private on-premise clusters, dedicated cloud VPCs (AWS/GCP) or a managed enterprise cloud instance | — | — | WebAssembly | 2014 | Site claims under 1 ms LongCell cold start and under 2 MB baseline memorySite claims under 1 ms LongCell cold start and under 2 MB baseline memory; runs TypeScript, JavaScript, WebAssembly (from C/C++/Go/Rust) and native Rust plugins; LongGuard signed-capability 'supply-chain firewall'; pitched for untrusted AI-generated code; part of i2c's larger 7-layer enterprise platform; no public docs, API reference, SDK or source code found (GitHub org i2cvn has only unrelated forks); site is a JavaScript single-page app |
| Core | Core | $0.017 | $0 compute when paused or stopped; disk included up to plan quota$0 compute when paused or stopped; disk included up to plan quota, then $0.08/GiB-month | Suspend keeps memory; auto-pause after idle (Free fixed at 30 min, paid 5-30 min or off)Suspend keeps memory; auto-pause after idle (Free fixed at 30 min, paid 5-30 min or off); stop keeps disks; Free runs capped at 120 min per run, growing to 480 min | VM | Oct 2026 | Default 4 vCPU/8 GiB with 20 GiB persistent /data; templates agent-browser (desktop + Chromium), android (Android 14)Default 4 vCPU/8 GiB with 20 GiB persistent /data; templates agent-browser (desktop + Chromium), android (Android 14), omarchy (Arch + Hyprland), ci; macOS machines; NVIDIA L4 GPUs; fork in seconds; snapshots; SSH; public HTTPS ports; drives Claude Code or Codex from the CLI; devcontainer support; TypeScript SDK, CLI, REST; sizes up to 32 vCPU/64 GiB on Team |
| Core | Core | $0.1656 | $0 compute ('No idle charges'); volumes $0.0002/GB-hour after 10 GB | Session VM reused across execute() calls, destroyed when the session endsSession VM reused across execute() calls, destroyed when the session ends; snapshots and volumes keep state; 'No idle charges' | VM | Aug 2025 | Sub-200 ms cold start (185 ms P95 claimed), snapshot restore <500 ms; snapshots, clone and fan-out; any OCI imageSub-200 ms cold start (185 ms P95 claimed), snapshot restore <500 ms; snapshots, clone and fan-out; any OCI image; persistent volumes; egress allowlists; secrets injection; public URL shares; computer use desktop with browser and human takeover; SSH-native CLI (ssh instavm.dev); Python and TypeScript SDKs; webhooks; open-source local coderunner |
| Core | Core | $0.30 | $0 ('No charge when stopped') | Lifecycle policy: pause_after_idle or pause_after (state snapshotted, compute releasedLifecycle policy: pause_after_idle or pause_after (state snapshotted, compute released, filesystem kept), delete_after from creation, auto_resume on_activity; manual pause, stop, resume | VM | Feb 2026 | Persistent Islo Machines that boot your stack (Postgres, Redis, browser); snapshots; terminal/SSH attachPersistent Islo Machines that boot your stack (Postgres, Redis, browser); snapshots; terminal/SSH attach; gateway injects credentials at egress, egress policies, per-request audit; GitHub, Linear, Jira, Slack integrations; TypeScript and Python SDKs, CLI; BYOC in your VPC; site footer says (c) Incredibuild, 'built by the team behind Incredibuild' |
| Core | Core | $0.11 | $0 while hibernated | Hibernate() pauses and persists memory, processes, open FDs and diskHibernate() pauses and persists memory, processes, open FDs and disk, resume() restores them (new guest IP, outbound TCP reset); hibernated sandboxes are not billed; auto-destroy idle timeout default 300 s, configurable via timeoutSec or setTimeout() | VM | Jun 2026 | Claims ready in 10 ms; any OCI image (Docker Hub, GHCR, private registries); fork running VMsClaims ready in 10 ms; any OCI image (Docker Hub, GHCR, private registries); fork running VMs; checkpoints and rollback; persistent shells with detach; public HTTPS URLs per port (WebSockets, SSE, HTTP/3); default-deny egress profiles enforced in the host kernel; credentials injected by an out-of-guest proxy; HMAC-chained audit log; browser and computer use; TypeScript SDK (MIT, github.com/isorun/sdk), MCP server; integrations with Claude Managed Agents, OpenAI Agents, LangChain, CrewAI; default 4 GiB scratch disk; SOC 2 Type II 'soon' |
| Core | Core | $0.1008 | Stopped: disk only, $0.101/GiB-monthStopped: disk only, $0.101/GiB-month; an idle running machine bills only the CPU time and memory it uses (site: about 14 cents a day) | Machines are persistent by default and keep running until stoppedMachines are persistent by default and keep running until stopped; stop keeps the disk (only disk bills), start restores it as it was; lifetime 'ephemeral' removes the machine when the SDK scope ends; ttl removes it at a deadline; snapshots and forks; when funds run out machines stop at once and disks are kept 24 h | VM | — | No instance types: one large machine size, boot with 2-64 vCPUs to pack hostsNo instance types: one large machine size, boot with 2-64 vCPUs to pack hosts; x86_64 or arm64 (EPYC 5, Xeon 6, Graviton 5); TypeScript, Python and Rust SDKs (@indexable/sdk), CLI and API; remote MCP server at ix.dev/mcp as a Claude connector; streaming exec; snapshots and forks; secrets and networking; spending ceilings per API key; by Indexable, Inc. (ix.dev's homepage now redirects to its blaze Rust build tool) |
| Core | Core | Flat per VM: Starter 1 vCPU/2 GiB/10 GiB $25/mo or $250/yrFlat per VM: Starter 1 vCPU/2 GiB/10 GiB $25/mo or $250/yr; Pro 2 vCPU/6 GiB/50 GiB $49/mo or $490/yr; Max 4 vCPU/12 GiB/100 GiB $99/mo or $990/yr (daily snapshot included); hourly at Starter size only: On-Demand $0.05/h, Spot $0.025/h, per second with a 60 s minimum, from prepaid credits ($10 minimum); private template slots $5-$29/mo; no free trial, 30-day money-back on the first plan; no 2 vCPU/4 GiB size | Flat plans: same as running (flat monthly fee)Flat plans: same as running (flat monthly fee); hourly VMs: $0 while stopped ("a stopped agent burns nothing") | Flat plans are always on and never auto-suspended for being idleFlat plans are always on and never auto-suspended for being idle; hourly VMs (On-Demand, Spot) bill per second while running and can be stopped anytime; Spot can be interrupted (an AWS reclaim can lose up to the last ~10 minutes of unsynced work; jurniti-initiated pauses lose nothing and restore automatically); /persist volume survives restart, resize and snapshot; after cancelling, the volume is kept to the end of the paid period, then deleted with no backups | VM | Jun 2026 | Managed hosting for 30 open-source agent harnesses (Claude Code, Claude Desktop, Codex CLI, OpenCode, GooseManaged hosting for 30 open-source agent harnesses (Claude Code, Claude Desktop, Codex CLI, OpenCode, Goose, Gemini CLI, Aider, Cursor Agent, Hermes, OpenClaw and others), one microVM per agent; bring your own model key, kept inside the VM, no token markup; public TLS subdomain <you>.jurniti.com; browser terminal; CLI, MCP server and REST API with OpenAPI spec; jurniti run fans out a fleet of workers and collects one result file per agent; snapshot a VM into a public or private template and fork it; jurniti transfer moves local agent sessions into the VM; homepage demo shows a microVM booting in 124 ms; a paid VM is live in about 3 minutes; region us-west-2 only |
| Core | Core | $0.0682 | $0 compute; stored data $0.05/GiB-month | One-shot sandboxes are destroyed after the runOne-shot sandboxes are destroyed after the run; held sandboxes expire 5 minutes after their last command (ttl up to 1 hour, one command capped at 10 minutes) and live in the host's memory; coding-agent workspaces sleep after ten quiet minutes by default and wake on the next request; apps scale to zero by restoring VM snapshots taken after startup; a paused or scaled-to-zero machine accrues only stored data | VM | Jul 2026 | Global app platform: static sites, HTML apps, shebang handlers, container services from Dockerfiles or ComposeGlobal app platform: static sites, HTML apps, shebang handlers, container services from Dockerfiles or Compose, a replicated SQLite database and file tree per app; sandboxes: kedge eval runs stdin in a fresh VM (Python 3.12, Node 20, Ruby 3.3, Bash, Deno; 256 MiB and 30 s by default), held sandboxes through a REST API with exec, optional no-network mode, no ingress; persistent coding-agent workspaces with Kedge's own agent, Claude Code or Codex; GitHub Actions runners; forkable VM snapshots (Show HN title); scale to zero |
| Core | Core | $0.1576 | $0 compute while asleep; snapshot $0.08/GB-month | Local: runs until stopped, agent-created sandboxes stop after 30 idle minLocal: runs until stopped, agent-created sandboxes stop after 30 idle min; save the running machine in ~0.7 s, rewind in ~1 s, live fork. kosa8 Cloud: idle sandbox sleeps on your timer, keeps a snapshot, next command wakes it | VM | — | Local sandboxes on Apple-silicon Macs, x86-64 Linux, Windows 11 via WSL2 (preview); kosa8 Cloud in Dallas and HelsinkiLocal sandboxes on Apple-silicon Macs, x86-64 Linux, Windows 11 via WSL2 (preview); kosa8 Cloud in Dallas and Helsinki; own hosts (BYOC); air-gapped Enterprise; checkpoint, rewind, fork; Docker-compatible commands; MCP server; Claude Code plugin; human approval gate; hash-chained audit log; keys and egress held on the host; 'ready in 1.9 ms' claim; 20 Cloud sandboxes at once per org; commercial, not open source |
| Core | Core | Billed as Koyeb Instances per second: Standard nano 0.25 vCPU/256MB $0.0036/hrBilled as Koyeb Instances per second: Standard nano 0.25 vCPU/256MB $0.0036/hr, micro 0.5/512MB $0.0072/hr (SDK default), small 1/1GB $0.0144/hr, medium 2 vCPU/2GB $0.0288/hr, large 4 vCPU/4GB $0.0576/hr; Eco eco-large 2 vCPU/4GB $0.0288/hr ($21.43/mo) exists but docs don't say sandboxes can use Eco (scale-to-zero is documented for Standard and GPU only); GPU instances from $0.50/hr; needs Starter, Pro ($29/mo + compute) or Scale ($299/mo + compute) plan | $0 compute in Deep Sleep (scaled to zero); Light Sleep billing not stated | Scales to zero after idle_timeout (default 300 s) into Deep Sleep (1-5 s cold start)Scales to zero after idle_timeout (default 300 s) into Deep Sleep (1-5 s cold start); optional Light Sleep snapshot (~200 ms wake, public preview); auto-delete after delete_after_delay or delete_after_inactivity_delay (60 s to 24 h, or 12 h after scaling to zero) | VM | Nov 2025 | Public preview; Python and JavaScript SDKs; any Docker image; file ops, background processes, expose ports, TCP proxyPublic preview; Python and JavaScript SDKs; any Docker image; file ops, background processes, expose ports, TCP proxy; snapshots in SDK; CPU and GPU sandboxes; multi-region; 'cold starts as low as 250ms'; Koyeb is joining Mistral AI (site banner) |
| Core | Core | $0.012 | $0 compute when powered off; disk $0.00005/GB-hr on full disk size (about$0 compute when powered off; disk $0.00005/GB-hr on full disk size (about $0.0365/GB-month) | No auto-suspend stated; power off stops vCPU and RAM billing at once and keeps the diskNo auto-suspend stated; power off stops vCPU and RAM billing at once and keeps the disk (billed on full size); start again in seconds; auto powers off when prepaid credit runs out; automatic snapshots and backups | VM | Apr 2019 | Boots in seconds; full root, systemd, Docker; SSH; live resize of vCPU, RAM and diskBoots in seconds; full root, systemd, Docker; SSH; live resize of vCPU, RAM and disk; RAM and disk reserved 1:1, never oversold; domains with automatic HTTPS; snapshots, clone and export; open-source SDK, CLI, MCP server and REST API; solution pages for AI-agent sandboxes, untrusted user code and CI runners; bare metal in Los Angeles only; also sells Nest managed web hosting |
| Core | Core | $0.189 | $0 compute while paused (GA blog: idle sandboxes pause so you don't pay)$0 compute while paused (GA blog: idle sandboxes pause so you don't pay); storage 0.000123 LSU/GiB-hr | Ephemeral by default; inactive TTL shuts idle sandboxes down / pauses them automaticallyEphemeral by default; inactive TTL shuts idle sandboxes down / pauses them automatically, hard TTL caps total runtime; snapshots capture full state for restore or copy-on-write forks | VM | May 2026 | Prewarming keeps p50 start under 0.98 s; full root, Docker inside; bring your own Docker image or templatesPrewarming keeps p50 start under 0.98 s; full root, Docker inside; bring your own Docker image or templates; Auth Proxy injects credentials into outbound requests; snapshots and copy-on-write forks; Blueprints; Service URLs and port tunneling; Sandbox CLI with ssh/scp/rsync/sftp; Python and TypeScript SDKs; works with Deep Agents or any framework; all LangSmith Cloud regions, BYOC and self-hosted; restricted egress allowlist for non-Enterprise orgs |
| Core | Core | $0.1656 | Paused state billing not published; free during preview | Timeout 1 min to 8 h (default 5 min); auto_pause at timeout saves memory and disk andTimeout 1 min to 8 h (default 5 min); auto_pause at timeout saves memory and disk and resumes on the next request; stop keeps disk only; delete removes | VM | Mar 2026 | ~100 ms start; snapshots of memory and disk, fork into parallel sandboxes~100 ms start; snapshots of memory and disk, fork into parallel sandboxes; templates from any container image (ECR, GCP AR, ACR); XFCE desktop 1440x900 for computer use; SSH and VNC; LSP for Python/TypeScript; firewall with domain/CIDR allowlists and header transforms for credential brokering; Python and TS SDKs; MCP server; 1-8 vCPU, 0.5-8 GB, 10 GiB disk; US only; BYOC or self-hosted on request |
| Core | Core | No sandbox rates on lightning.ai/pricing or the sandbox docsNo sandbox rates on lightning.ai/pricing or the sandbox docs. Instance types: cpu-1 (1 vCPU / 1 GB, default), cpu-2 (2 vCPU / 8 GB), cpu-4 (4 / 16 GB), cpu-8 (8 / 32 GB), cpu-16 (16 / 64 GB); spot option. No 2 vCPU / 4 GiB type | — | Non-persistent sandboxes are deleted on stop()Non-persistent sandboxes are deleted on stop(); with persistent=True, stop() auto-snapshots the filesystem and resume() brings it back; timeout sets a maximum lifetime in ms (0 = none) and can be extended; manual snapshots restore into new sandboxes | — | May 2025 | Python SDK (lightning-sdk), JS SDK (@lightningai/sdk) and 'lightning sandbox' CLIPython SDK (lightning-sdk), JS SDK (@lightningai/sdk) and 'lightning sandbox' CLI; runtimes node24, node22, python313 or your own OCI image; sudo in every sandbox; 'start in milliseconds' claim; snapshots; PTY sessions; detached long-running commands; preview URLs for dev servers; egress network_policy (allow-all, deny-all or custom rules); GPU instance types exist (custom images are CPU-only); Lightning AI runs on AWS and GCP |
| Core | Core | Not published; public preview launched 2026-10-01 with two weeks of free usageNot published; public preview launched 2026-10-01 with two weeks of free usage; the pricing page covers only LithosAI's inference API | — | Snapshot, branch and restore a sandbox's full running state (filesystemSnapshot, branch and restore a sandbox's full running state (filesystem, processes and memory); the launch post says environments can sleep and resume quickly while keeping running state; no timeouts or limits published | VM | Oct 2026 | Own benchmark: 50 ms new sandbox + first command, 95 ms snapshot with 1 GB of dirty dataOwn benchmark: 50 ms new sandbox + first command, 95 ms snapshot with 1 GB of dirty data, 322 ms branch and 140 ms restore of a 512 MB environment, with running processes restored; launched alongside LithosAI's low-latency inference (Lithos API serving Kimi, GLM, DeepSeek and Qwen models); sign-up at console.lithosai.cloud; public docs cover only the inference API |
| Core | Core | $0.0162 | $0 compute when paused; $0.14 per used GB-month of disk | Running sandbox billed even when idle; pause keeps files only (memory and processes endRunning sandbox billed even when idle; pause keeps files only (memory and processes end, resume boots fresh); fork copies a running sandbox with memory; default lifetime 5 min from SDK/CLI, none from API/dashboard; a sandbox with no deadline stops after 596 h (~24 days); trial sandboxes deleted after 1 h | VM | Jun 2026 | 228 ms median start claim (2026-10-12, 300 runs, Small)228 ms median start claim (2026-10-12, 300 runs, Small); templates base, interpreter, codex, claude, opencode, pi, prime, desktop, desktop-camoufox; desktop + Chromium; fork with memory; snapshots and restore; Persistent Volumes at /workspace up to 100 GB; private preview URLs; no SSH protocol (CLI shell only); no GPU; no self-hosting; US East and EU West; TypeScript and Python SDKs, CLI, HTTP API; Pro 40 vCPU concurrent, Pro 5x 220, Pro 20x 1,200; also sells app hosting, Postgres, Redis, object storage |
| Core | Core | $0.052 | Stopped: same as running (resources reserved)Stopped: same as running (resources reserved); suspended: $0 compute, snapshot $0.078/GB-month | VMs stay on until stopped or suspended; stop keeps the cloud instance and bills at fullVMs stay on until stopped or suspended; stop keeps the cloud instance and bills at full rate; suspend snapshots the disk, deletes the instance and stops machine billing (snapshot billed as an image), resume with machine0 start gets a new IP; opt-in idle detection suspends or destroys after 30 min of inactivity by default; machines snapshotted and destroyed if the wallet runs out | VM | Jun 2026 | Persistent VMs designed for long-running agents; NixOS or Ubuntu images (open source) with Docker, Node, PythonPersistent VMs designed for long-running agents; NixOS or Ubuntu images (open source) with Docker, Node, Python, Claude Code and Codex preinstalled; every operation is a CLI command with --json, plus a remote MCP server; profiles inject credentials, MCP servers and prompts for Claude Code and Codex; golden images and clones; static public IP and HTTPS at <vm>.mac0.io; in-browser shell; 9 regions; GPUs H100, H200, L40S, MI300X; site says 99.99% uptime |
| Core | Core | $0.022 | Agent box $1/mo per 2 GB awake or asleepAgent box $1/mo per 2 GB awake or asleep; sleeping sandbox, desktop or phone $0 compute + $0.08/GB-month disk | Agent boxes sleep when idle and wake on the next message in about 1 s (0.6 s median)Agent boxes sleep when idle and wake on the next message in about 1 s (0.6 s median), disk kept; sandboxes end with their disk deleted when the job is done unless set to pause when idle; desktops and phones sleep with disk kept (Windows wakes in about 1 min) | VM | Jul 2026 | Wake in ~1 s; Linux and Windows Server 2022 desktops, Android phones; swarms up to 10,000 sandboxesWake in ~1 s; Linux and Windows Server 2022 desktops, Android phones; swarms up to 10,000 sandboxes; Telegram, WhatsApp, Slack, email channels; model gateway at list price; US or Germany; HIPAA BAA on request; TypeScript and Python SDKs, CLI, REST, MCP; harnesses OpenClaw, Hermes, Claude Agent SDK, Codex or own Docker image; private networks, warm pools; startup credits up to $10,000 |
| Core | Core | Starter $20/mo (one agent or task at a time, always-on computer, limited storage)Starter $20/mo (one agent or task at a time, always-on computer, limited storage); Builder $100/mo (three parallel agents, scheduled tasks, region selection); Max $200/mo (six to eight parallel agents, maximum compute, multiple workspaces, static IP and custom images); 14-day trial with card; AI subscriptions billed by your provider; CPU and RAM per plan not stated; Enterprise custom | Same as running (flat monthly) | Always-on computers on every hosted plan; snapshot history limited on StarterAlways-on computers on every hosted plan; snapshot history limited on Starter, 7 days on Builder, 30 days on Max | VM | Jul 2026 | 'EU-first AI workspace': a cloud computer with chat, terminal, files, apps and a Postgres database'EU-first AI workspace': a cloud computer with chat, terminal, files, apps and a Postgres database, used through the Matrix Desktop app, a browser or the CLI; runs Claude Code, Codex, Cursor, OpenCode, Pi, Gemini CLI or other terminal agents on your own provider accounts, plus Hermes and OpenClaw-style agents; Symphony for parallel coding agents; integrations (GitHub, Google Drive, Gmail, Calendar, Slack); managed cloud or self-host preview on your own VPS; AGPL-3.0; 88 stars; mobile apps in early access; AI Gateway on a waitlist |
| Core | Core | $0.108 | $0 compute when stopped; retained root disk $0.0000625/GiB-hour$0 compute when stopped; retained root disk $0.0000625/GiB-hour; volumes $0.15/GiB-month; snapshots $0.08/GiB-month | Stop ends CPU and RAM billing; retained root disks keep billing until deletedStop ends CPU and RAM billing; retained root disks keep billing until deleted; ephemeral or persistent sandboxes on paid plans; free trial sessions limited to 1 hour; snapshots save disk state | VM | Oct 2024 | Open source Apache-2.0 (github.com/superradcompany/microsandbox, 9.1k stars)Open source Apache-2.0 (github.com/superradcompany/microsandbox, 9.1k stars); local runtime with no daemon or root, boot under 100 ms; cloud in private beta by request; BYOC; Builder up to 16 vCPU / 16 GiB, Teams up to 32 vCPU / 32 GiB, 64 GiB root disk; credential placeholder substitution; network egress policy; SDKs in Rust, TypeScript, Python, Go and Ruby; runs on macOS, Linux and Windows |
| Core | Core | $0.0893 | $0 compute when paused or stopped; disk at the snapshot rate $0.018/GB-month$0 compute when paused or stopped; disk at the snapshot rate $0.018/GB-month when paused (FAQ); attached disk $0.09/GB-month while the machine exists | Sandboxes go idle after 30 minutes by default (accounts opened since 2026-10-11Sandboxes go idle after 30 minutes by default (accounts opened since 2026-10-11; idle_timeout_sec 0 disables); pause keeps memory and processes for at most 30 minutes, then behaves like a stop; stop keeps the disk and wakes on the next request; machines pause at a $0 balance | VM | — | Claims 57.1 ms sandbox startup (ComputeSDK benchmark, 2026-10-23)Claims 57.1 ms sandbox startup (ComputeSDK benchmark, 2026-10-23); persistent desktop computers with screenshot, click and type APIs; deployments, Postgres/MySQL/Redis/MongoDB, custom domains; fork, checkpoints, snapshots, volumes; live vCPU/memory resize; SDKs TypeScript, Python, Go, Rust, Elixir, Java; CLI and MCP; BYOC in AWS or Google Cloud; white-label workspaces; limits up to 16 vCPU/32 GB per machine on Enterprise |
| Core | Core | $0.238 | — | — | — | — | 2 vCPU = 1 physical core in Modal pricing |
| Core | Core | $0.0877 | Flat monthly price; no lower price for a stopped VM stated | Always-on monthly VMs: the first month is charged at creation and renews monthly from aAlways-on monthly VMs: the first month is charged at creation and renews monthly from a prepaid balance; stop, start, resize and rebuild; if the balance can't cover egress the VM is stopped until a top-up; a short balance starts a grace period before suspension | VM | Feb 2026 | Hosting arm of an agent social network (identity, email inbox, Coinbase CDP wallet on Base, marketplace)Hosting arm of an agent social network (identity, email inbox, Coinbase CDP wallet on Base, marketplace); an agent's own API key buys a VM, database and hostname with no human account; mbd CLI, REST API and MCP server (157 tools); free name.moltbotden.com subdomain; object storage in us-central1; crypto mining banned (a shared-core VM at full CPU for 6 h is stopped for review) |
| Core | Core | Per agent, always on, prorated daily: Small $2.88/mo (up to 1.5 vCPU, 1.5 GB, 5 GB NVMe)Per agent, always on, prorated daily: Small $2.88/mo (up to 1.5 vCPU, 1.5 GB, 5 GB NVMe); Large $5.25/mo (up to 3 vCPU, 3 GB, 10 GB); XLarge $7.90/mo (up to 3 vCPU, 6 GB, 20 GB); yearly billing two months free; price built from $0.70/vCPU-month + $0.65/GB-month RAM + $0.07/GB-month NVMe + $0.50 flat per agent, but other sizes only via Enterprise; no 2 vCPU / 4 GB size at sign-up; no network fees; paid tools (cloud browser, integrations, memory, voice) debited from a prepaid wallet; Enterprise and on-premise quoted | Stopped agent: 10x cheaper than runningStopped agent: 10x cheaper than running, workspace kept (exact rate not stated); deleted: $0 | Always on, never put to sleep; a stopped agent keeps its workspace and costs 10x lessAlways on, never put to sleep; a stopped agent keeps its workspace and costs 10x less; deleted agents cost $0; billing prorated to the day the agent exists; hot resize without restart | — | — | Hosts Mue (their agent), OpenClaw, Hermes, Claude Code, Codex or your own harnessHosts Mue (their agent), OpenClaw, Hermes, Claude Code, Codex or your own harness; OpenClaw cold start 1.1 s, provisioning under 7 s, '2x faster per core than a cloud VM' (internal measurements); automatic repair; versioned workspace and file explorer; browser console, SSH, API and public HTTPS URL per agent; molted-mcp with cloud browser and 1,500+ integrations; white-label agent platforms for integrators and hosting resellers; regions EU and North America; bring your own model key |
| Core | Core | $0.10 | $0 compute while paused; storage still billed (snapshot storage 1 MCU per 5$0 compute while paused; storage still billed (snapshot storage 1 MCU per 5 TB-hour = $0.00001/GB-hr) | Pause keeps memory and processes (no compute charge, storage still billed)Pause keeps memory and processes (no compute charge, storage still billed); stop for one-off tasks; TTL with a pause or stop action; wake-on-request resumes a paused instance on HTTP or SSH access; snapshot and branch at any time | VM | Nov 2024 | Infinibranch: snapshot, branch and replicate running VMs; Liquid Metal burst scaling (blog: p90 start 45 s to 3.96 s)Infinibranch: snapshot, branch and replicate running VMs; Liquid Metal burst scaling (blog: p90 start 45 s to 3.96 s); Devboxes with VS Code over SSH, tmux, live previews, scale-to-zero; MorphVM limits 64 vCPU/256 GB/1 TB (Free) up to 1024 vCPU/4 TB RAM (Scale); devbox concurrency 8/32/128 by tier; GitHub Actions runners; Python and TypeScript SDKs |
Moru CoreHosted service appears shut down: moru.io now shows 'MoruHosted service appears shut down: moru.io now shows 'Moru, an AI-native game studio from Korea' and /docs and /pricing return 404 | Core | Moru.io/pricing now returns 404; the code is open source (Apache-2.0) and the runtime canMoru.io/pricing now returns 404; the code is open source (Apache-2.0) and the runtime can be self-hosted from moru-ai/sandbox-infra | — | Templates are paused microVM snapshots that new sandboxes resume fromTemplates are paused microVM snapshots that new sandboxes resume from; no idle policy stated in the README; the docs at moru.io/docs now return 404 | VM | Jan 2026 | Cloud runtime for agent harnesses such as Claude Code or Codex, one microVM per sessionCloud runtime for agent harnesses such as Claude Code or Codex, one microVM per session; forked from E2B (SDK and CLI) and E2B infra; TypeScript (@moru-ai/core) and Python (moru) SDKs and a CLI; custom templates from a Dockerfile with CPU and memory limits; memory snapshots lazy-loaded through userfaultfd so sandboxes start within a second; 78 GitHub stars (moru) and 11 (sandbox-infra) |
| Core | Core | $0.05 | $0 while hibernated | Automatic pause to a snapshot on local NVMe when the agent waitsAutomatic pause to a snapshot on local NVMe when the agent waits, resume on the next request; hibernated sandboxes cost $0 and count toward the 1,000-sandbox limit; mos run / task mode destroys the sandbox after the command | VM | — | REST API, mos CLI, Python/TypeScript/Go SDKs, MCP and native SSH; task mode clones a repo, installs and runs a commandREST API, mos CLI, Python/TypeScript/Go SDKs, MCP and native SSH; task mode clones a repo, installs and runs a command; lifecycle mode with /workspace and durable processes; forks; one-shot run ~1.0 s and warm exec ~280 ms (own benchmark, Aug 2026); egress allowlists and deny-all; workload identity via OIDC for AWS/GCP/Vault; RBAC, SSO, audit logs, webhooks, OpenTelemetry; agent configs for Codex, Claude, Cursor; team says it helped build infrastructure at Cloudflare |
| Core | Core | $0.018 | Persistent paused sandbox: saved state $0.0002/GiB-hourPersistent paused sandbox: saved state $0.0002/GiB-hour; non-persistent sandboxes cannot idle cheaply (full price until timeout) | No idle suspend: a sandbox pays its full size idle or busy until its timeout (10 s to 24No idle suspend: a sandbox pays its full size idle or busy until its timeout (10 s to 24 h, default 1 h); persistent sandboxes pause at timeout or on request, keeping memory, processes and disk; paused sandboxes deleted 30 days after the pause | VM | Nov 2026 | Starts in about a second (presets restore from snapshots); pause and fork; desktop template; SSH and public portsStarts in about a second (presets restore from snapshots); pause and fork; desktop template; SSH and public ports; private sandbox networks; templates base, ubuntu with Docker, agent (Claude Code, Codex, OpenCode, DeepSeek Harness, pi, OpenClaw), desktop; TypeScript SDK/CLI (@munitoai/cli); EU and US regions; 32,752 sandboxes in one run; also sells open-model inference in the EU |
| Core | Core | $0.1656 | $0 (paused: no vCPU, memory or storage billed) | Paused after a few idle minutes and resumed when a session needs itPaused after a few idle minutes and resumed when a session needs it; paused computers keep their disk and bill nothing; any command, file operation or snapshot resumes it; auto-destroyed at a fixed 12-hour max runtime; checkpoints and forks from a running VM; no persistent volumes | VM | Aug 2026 | Main product is 'autonomous companies' run by agent teams on Naive Managed AgentsMain product is 'autonomous companies' run by agent teams on Naive Managed Agents; computers are also created directly through the API, SDK and CLI (naive computer create): 1-16 vCPU, 128 MiB-64 GiB memory, up to 100 GiB disk, default 2 vCPU / 4 GiB (session computers default 1 vCPU / 1 GiB); optional managed browser with saved logins in a vault; cloud Android phones and iPhones on residential IPs; no cap on concurrent computers; docs and API also served from vetta.sh |
| Core | Core | No sandbox rates published; beta by access requestPricing page | No running state between operations; image storage price not published | No long-lived sandbox: each operation starts fresh compute from an imageNo long-lived sandbox: each operation starts fresh compute from an image, runs one command and ends; changed files can be saved as an image (filesystem only, no process memory); checkpoint images kept 180 days, then untagged unreferenced images may be deleted | VM | — | Part of Nebius Token Factory; Python SDK, REST API, CLI and MCP (Contree SDK/CLI/MCP)Part of Nebius Token Factory; Python SDK, REST API, CLI and MCP (Contree SDK/CLI/MCP); images from setup commands, OCI import or Dockerfile; branch from checkpoints; subprocesses with stdin; SSE output streaming; network on by default, can start with no guest network interface; 50 simultaneous operations during beta; guide for SWE agents |
| Core | Core | Free tier: 1,500 sandbox minutes per organisation per month (25 hours at 2 vCPU / 4 GB orFree tier: 1,500 sandbox minutes per organisation per month (25 hours at 2 vCPU / 4 GB or 50 hours at 1 vCPU / 2 GB) and 20 GB free disk shared across sandboxes; usage beyond that billed from a prepaid wallet at rates not published in the docs; KYC required | $0 compute while paused; disk beyond the 20 GB free allowance billed (rate not$0 compute while paused; disk beyond the 20 GB free allowance billed (rate not stated) | Auto-pause after an idle window (default 15 minutesAuto-pause after an idle window (default 15 minutes; idle_timeout_seconds, 0 turns it off): process memory and filesystem are checkpointed and compute billing stops; the next API or MCP call wakes it with processes intact; on_idle can be 'delete' instead; optional max_lifetime_seconds (none by default); paused sandboxes deleted after 14 days by default (paused_retention_seconds); snapshots with undo and duplicate (fork) | Process | — | Templates Ubuntu 26.04 dev (default), Docker, Python interpreter and Playwright/Chromium desktop with VNC (site alsoTemplates Ubuntu 26.04 dev (default), Docker, Python interpreter and Playwright/Chromium desktop with VNC (site also lists Debian 12 and Ubuntu 24.04 minimal); configurable CPU, memory and storage; egress deny_all by default with allow-lists; preview URLs; audit trail of commands; Sandbox MCP server; JavaScript and Python SDKs; ComputeSDK provider; tutorials for Claude Code, Codex and OpenClaw; India-resident sovereign cloud (MeitY empanelled, SOC 2 Type II) |
| Core | Core | $0.048 | $0 compute per product page; /workspace on ABS $0.00013/GB-hr$0 compute per product page; /workspace on ABS $0.00013/GB-hr ($0.0949/GB-month); cluster node billing while sandboxes are paused not stated | Pause releases compute and keeps /workspace files on ABS block storage, resume instantPause releases compute and keeps /workspace files on ABS block storage, resume instant; ephemeral sandboxes have no disk and leave nothing on delete; expiration can be renewed via API; creating from snapshots not supported yet | Container | — | Ephemeral boot ~2.5 s, persistent first boot 6-7 s; instant pause/resume/delete; OpenSandbox-compatible Lifecycle APIEphemeral boot ~2.5 s, persistent first boot 6-7 s; instant pause/resume/delete; OpenSandbox-compatible Lifecycle API; SDKs Python, TypeScript, Go, Java, Kotlin, C#; ports reached via signed expiring gateway endpoints; AMD EPYC 9354P + DDR5 on Nirvana's own cloud (no hyperscalers); region us-sva-2 (Silicon Valley), not every region; claims 20% cheaper than leading sandbox providers |
| Core | Core | $0.0357 | $0 compute; retained storage billed (rate not stated) | Stops after idleTimeout (default 5 min, 1 min to 24 h, or never) and keeps /workspaceStops after idleTimeout (default 5 min, 1 min to 24 h, or never) and keeps /workspace; onIdle Stop or Delete; maxLifetime default 24 h (up to 720 h); pause with process memory kept up to 30 min (beta, CPU only); maxCostUSD cap | Container | Sep 2026 | Part of a broader GPU compute platform (training, inference, jobs, functions, BYOCompute pools)Part of a broader GPU compute platform (training, inference, jobs, functions, BYOCompute pools); Python SDK (Modal-compatible import), CLI, HTTP API, MCP for Claude Code/Codex/Cursor; directory snapshots (beta); no outbound network unless opened; sizes 0.25 to 64 vCPU and up to 256 GiB per its E2B comparison |
| Core | Core | $0.0667 | — | Cost of paused or scaled-to-zero services not stated | VM | — | BYOC; GPU sandboxes (e.g. L4 $0.80/hr) |
| Core | Core | $0.1166 | $0 compute while paused; persistent storage free up to 60 GB per account$0 compute while paused; persistent storage free up to 60 GB per account, then $0.00009/GB-hour | Default timeout 5 min from creation, extendableDefault timeout 5 min from creation, extendable; session cap 1 h Free, 3 h Paid, 4 h Enterprise unless long-running mode (metadata long_running) lifts it; at timeout killed by default or paused (lifecycle.on_timeout=pause, auto_resume); optional idle timeout when no client is connected; pause keeps filesystem and memory including running processes, resume in about 1 s | — | Jul 2025 | Starts in under 200 ms on average (site claim); pause/resume with memoryStarts in under 200 ms on average (site claim); pause/resume with memory; snapshots, volumes, secrets, SSH access, interactive terminal; desktop sandboxes with VNC streaming for computer use; browser automation; Python and JavaScript/TypeScript SDKs (novita_sandbox, novita-sandbox) and CLI; limits Free 2 vCPU/4 GiB, 5 concurrent; Paid 8 vCPU/8 GiB, 100 concurrent; regions us-phx-1 (default) and legacy us-virginia-1 |
| Core | Core | $0.092 | Stopped: CPU and RAM released, retained disk $0.05/GiB-monthStopped: CPU and RAM released, retained disk $0.05/GiB-month; paused: RAM still billed ($0.008/GiB-hour), no CPU | Permanent VMs keep their files when stopped and can opt into 'Sleep when idle' (idle timePermanent VMs keep their files when stopped and can opt into 'Sleep when idle' (idle time of at least 2 minutes; memory saved to disk; wakes on an HTTP request to an exposed URL); pause freezes processes with memory kept; temporary VMs have a TTL that stops, pauses or removes them on expiry | VM | Mar 2026 | Linux VMs for agents with a Runtime API on port 9990 (exec with SSE streaming, files, ripgrep searchLinux VMs for agents with a Runtime API on port 9990 (exec with SSE streaming, files, ripgrep search, PTY over WebSocket, filesystem watch); TypeScript SDK and CLI; SSH; browser desktops; root; custom images; private networking between chosen VMs; HTTPS previews and custom domains; managed workloads with logs; disk archives; resource pools and customer namespaces for reselling VMs; inbound traffic blocked by default, reached through short-lived gateway tokens |
| Core | Core | No sandbox price found on Oracle's price list or docsNo sandbox price found on Oracle's price list or docs; shapes SMALL 1 OCPU / 4 GB, MEDIUM 2 OCPU / 8 GB, LARGE 4 OCPU / 16 GB | Expired sandboxes end; snapshot storage price not stated | Each sandbox has an expiration set at create (maximum initial duration 48 hours)Each sandbox has an expiration set at create (maximum initial duration 48 hours), extendable before it expires; the service ends execution at expiry; filesystem snapshots (separate expiry) seed new sandboxes but do not keep processes or memory; per-command timeouts; optional warm pools per project | Process | — | Part of OCI Generative AI / OCI Enterprise AI projects (OpenAI-compatible Responses API, Code InterpreterPart of OCI Generative AI / OCI Enterprise AI projects (OpenAI-compatible Responses API, Code Interpreter, hosted agent applications); create through the OCI CLI or SDK, upload files, run commands, download results; managed runtime Oracle Linux with Python 3.11 or a custom image; project-level outbound networking rules; regions Chicago, Osaka, Ashburn, Phoenix, Frankfurt, London, Sao Paulo, Hyderabad, Riyadh; an OpenAI Agents API sandbox provider |
| Core | Core | $1.10/hour per sandbox or CI task, billed per second, no minimum; $5 free credit$1.10/hour per sandbox or CI task, billed per second, no minimum; $5 free credit; sandboxes share an 18-core, 36 GB Mac Studio M5 Max with no fixed CPU or memory allocation | $0 after close (no paused state) | Sandbox lives 1 h by default, 24 h at most; closes after 10 idle minutes by defaultSandbox lives 1 h by default, 24 h at most; closes after 10 idle minutes by default; kill() closes it and stops billing; GUI sandbox recordings kept 7 days | Process | Nov 2026 | macOS 27, Xcode 27.2, iOS Simulator; headless or GUI sessions with screenshots and screen-agent runs (runAgent withmacOS 27, Xcode 27.2, iOS Simulator; headless or GUI sessions with screenshots and screen-agent runs (runAgent with your Anthropic key); GitHub Actions runner (runs-on: offstage-macos); per team 4 open sandboxes and 4 running CI jobs; HTTP API, OpenAPI spec, TypeScript SDK 0.3.1, agent SKILL.md; a product of Exla |
| Core | Core | Managed service: no public price list or self-serve planManaged service: no public price list or self-serve plan, limits and price agreed per team; self-host free (Apache-2.0 announced, source 'opens soon') | — | Short-lived: created by restoring a warm memory snapshotShort-lived: created by restoring a warm memory snapshot; stops when its timeout runs out; kill deletes the VM, disk snapshot and network namespace | VM | Mar 2026 | Create-run-teardown trace 0.71 s (restore 438 ms); e2b-style REST API; TypeScript and Python SDKs, CLICreate-run-teardown trace 0.71 s (restore 438 ms); e2b-style REST API; TypeScript and Python SDKs, CLI; templates python-3.11, node-20, claude-code or from a Docker image; offline by default, HTTPS allowlists; preview URLs; audit log, credential vault; managed hosts all in Finland; single-node self-host on x86_64 KVM; worker for Claude Managed Agents (AGPL); by A14A B.V. |
| Core | Core | Not published: no pricing page, 'Talk to us'Not published: no pricing page, 'Talk to us'; the enterprise tier adds VPC deployment and SOC 2 Type II; model usage is metered per cell and per customer | ~$0 compute while paused; snapshot storage price not published | Agents idle-pause by themselves: the instance is snapshotted to S3 and deleted (computeAgents idle-pause by themselves: the instance is snapshotted to S3 and deleted (compute and local disk), and the next request or start materialises a fresh instance from the snapshot (200 ms per the homepage, 'a few hundred milliseconds' per the docs); agent.sleep() parks an agent for days at ~$0; destroy removes the snapshots too | VM | Apr 2026 | Agent runtime plus evals: an agent is one TypeScript file (instructions, model, tools, budgetAgent runtime plus evals: an agent is one TypeScript file (instructions, model, tools, budget, evals) deployed through an API; cells can be snapshotted and forked (code, files and SQLite database cloned together); exec and long-running services routed to the agent's URL; base images such as oncell/node; per-customer logged-in Chromium browser cells; daily budgets enforced below the model; the run log doubles as trace and audit trail; customer stories JustCopy.ai and OpenCrew; SSO; site says the SDK is open source, but github.com/oncellai/oncell returns 404 (the account has only cli and control-plane repos) |
| Core | Core | $0.2376 | $0 compute when stopped; retained root disks$0 compute when stopped; retained root disks, snapshots and volumes billed as storage ($0.000072-$0.00012/GiB-h) | Stop ends compute billing and keeps recoverable sandbox stateStop ends compute billing and keeps recoverable sandbox state; archive keeps it for later restore; delete releases it; per-sandbox quotas include idleTimeoutSeconds (crash protection) and maxDurationSeconds; snapshots can be forked into new sandboxes | — | Jun 2026 | Company leads with model post-training (Ponder Training) and serverless inferenceCompany leads with model post-training (Ponder Training) and serverless inference; cloud sandboxes driven from server code through the openpond-sdk TypeScript SDK, CLI and MCP: commands, files, processes, HTTPS preview ports; durable Sandbox Runtimes with events, checkpoints and wait-for-user state; snapshots and forks; reservation budgets and spend/duration/command quotas; open-source harness, desktop app and SDK in openpond/openpond (MIT, 4 stars) |
| Core | Core | $0.017 | $0 (paused sandboxes are free) | With auto-pause on, a sandbox pauses after 5 idle minutes and the next call wakes it withWith auto-pause on, a sandbox pauses after 5 idle minutes and the next call wakes it with memory and files intact; no session time limit, a sandbox lives until deleted; paused sandboxes are kept | VM | 2022 | Drop-in for the E2B JS/Python SDKs and CLI (set E2B_DOMAIN=sandbox.openscale.so); root in each microVMDrop-in for the E2B JS/Python SDKs and CLI (set E2B_DOMAIN=sandbox.openscale.so); root in each microVM; 100 sandboxes at once by default; background commands, streaming and stdin, interactive terminal, files, preview URL per port, Dockerfile templates, headless browser; same API key as its open-model inference API |
| Core | Core | $0.015 | $0 compute; storage $0.05/GiB-month | Idle boxes suspend automatically and resume on SSH with processes intactIdle boxes suspend automatically and resume on SSH with processes intact; $0 while suspended | VM | Nov 2026 | Millisecond start (demo shows 0.16 s); plain SSH with a hostname per box; Linux desktop in the browserMillisecond start (demo shows 0.16 s); plain SSH with a hostname per box; Linux desktop in the browser; Claude Code, Codex and Pi preinstalled with Docker and Google Chrome |
| Core | Core | EUR only: EUR 0.045/vCPU-hour + EUR 0.015/GiB-hourEUR only: EUR 0.045/vCPU-hour + EUR 0.015/GiB-hour, per second on allocated size while running (busy or idle); sizes small 1 vCPU/1 GB ~EUR 0.06/h, medium 2 vCPU/4 GB ~EUR 0.15/h, large 4 vCPU/8 GB ~EUR 0.30/h (GB treated as GiB); storage EUR 0.10/GB-month after 10 GB free; trial credit EUR 5 (no card) or EUR 25 (card); no plan fee. 2 vCPU/4 GiB = EUR 0.15/h; no USD price published | $0 compute when paused; snapshot counts as storage$0 compute when paused; snapshot counts as storage, first 10 GB free then EUR 0.10/GB-month | Pause stops the compute meter and keeps a snapshot (memory included)Pause stops the compute meter and keeps a snapshot (memory included), resumable on another host; per-call timeout 10 s to 24 h, default 600 s, extendable; free tier capped at 1 h continuous running; paused snapshots kept 7 days (free), 30 days (paid), no expiry (higher tier) | VM | — | EU-only hardware in Falkenstein (DE), Helsinki (FI), Roubaix (FR), EU operatorEU-only hardware in Falkenstein (DE), Helsinki (FI), Roubaix (FR), EU operator; cold start ~150 ms, under 30 ms from a warm pool; snapshot and resume; custom templates boot in ~300 ms; network off by default, allowlist or open egress; public preview URLs; volumes up to 100 GB; Docker in sandboxes; Python and JS SDKs, CLI, MCP server; E2B-style API |
| Core | Core | $0.1728 | $0 compute when hibernated; hibernated state storage currently included | Hibernate writes state to disk and stops CPU and RAM meteringHibernate writes state to disk and stops CPU and RAM metering; apps and databases scale to zero; Free sandboxes deleted 1 h after creation; paid TTL is an idle timeout (Pro up to 7 days, Team 30 days) | VM | Apr 2026 | 49 ms p50 / 64 ms p99 boot from snapshot claim; ~400 ms same-host fork; snapshot, fork, resume49 ms p50 / 64 ms p99 boot from snapshot claim; ~400 ms same-host fork; snapshot, fork, resume; templates for code interpreter, browser (Chromium + Playwright), coding agent (claude, codex, opencode, amp, gemini, copilot CLIs), Claude Managed Agents; also Postgres 16, apps, functions, cron; Python and TypeScript SDKs; MCP server; open source Apache-2.0 (pandastack-io/pandastack-ai, 51 stars); self-host; BYO-VPC on Enterprise |
| Core | Core | $0.0001 per second for compute and browser time ($0.36/hour)$0.0001 per second for compute and browser time ($0.36/hour); AI models $0.02 per 1,000 Neurons; Agent Memory, preview URLs and file I/O included; $5 free credits for new accounts (~14 hours); top-ups $1, $5, $20 (+10% bonus), $100 (+20% bonus) via MPP (Tempo or Stripe). Container CPU and memory not stated | $0 (only the included 10 MB Agent Memory persists) | Sandbox files are ephemeral and lost when the session endsSandbox files are ephemeral and lost when the session ends; 10 MB Agent Memory persists across sessions; background processes started over HTTP auto-terminate after 15 minutes (use the CLI's 'ppod start' for longer); command timeout defaults to 30 s | Container | Jan 2026 | CLI (@paperpod/cli, 'ppod') plus HTTP and WebSocket API; Python, JavaScript (node or bun) and shellCLI (@paperpod/cli, 'ppod') plus HTTP and WebSocket API; Python, JavaScript (node or bun) and shell; stable public preview URLs on paperpod.work; headless Chrome/Playwright screenshots, PDFs, scraping and test assertions; 50+ AI models (Llama, Qwen, Mistral, FLUX, Whisper); preinstalled ffmpeg, imagemagick, pandoc, sqlite3, git, gh, ripgrep; outbound network allowed; email magic-link sign-up; SKILL.md for agents |
| Core | Core | $0.1656 | $0 compute while paused; storage $0.15/GB-month | Pause writes memory and running processes to disk and frees RAM, resume ~120 ms p50Pause writes memory and running processes to disk and frees RAM, resume ~120 ms p50; idle sandboxes pause themselves; paused sandboxes bill storage only; Free sandboxes are leased and end after 4 h; Pro 24 h sessions + pause/resume; paid plans add a persistent workstation that pauses when idle; fork and snapshots | VM | Sep 2026 | microVM boot 110 ms, pause-to-resume 120 ms (their production measurements, July and November 2026)microVM boot 110 ms, pause-to-resume 120 ms (their production measurements, July and November 2026); fork a running sandbox with processes; snapshots; 25+ templates; 73 MCP tools (42 in open source); Python and TypeScript SDKs, CLI; preview URLs, custom domains, TCP, private networks; agents via Telegram, cron, functions, CI; self-host free with one install script; open source Apache-2.0 (github.com/pilldrops-oss/pilldrops, 0 stars) |
| Core | Core | No published prices; access by requestPricing page | — | Pause keeps the running environment until its expiry (for projects where pause isPause keeps the running environment until its expiry (for projects where pause is enabled); Stop shuts down compute and a later Start restores saved account state and skills; artifacts persist while stopped; download other files before Stop | gVisor | Oct 2026 | Persistent cloud computer for agents: browser, shell, files and a desktop a person can take over (Web SDKPersistent cloud computer for agents: browser, shell, files and a desktop a person can take over (Web SDK @pinesandbox/computer-web); one Computer with many sessions sharing its sign-ins; Pine's built-in agent or your own; skills learned from completed runs; server SDKs for Ruby, Go, TypeScript and Python; saved state encrypted with the customer's key under TPM; Pine-reported 78.3% checkpoint score on SaaS-Bench with GPT-5.6 Luna; starter repo PineComputer/pine-computer-starter; by Pine AI |
| Core | Core | Billed per second; no published ratesPricing page | — | Stop and resume (idle behaviour not stated) | VM | — | About 400 ms boot; snapshot and fork; OAuth 2.1 or scoped API keyAbout 400 ms boot; snapshot and fork; OAuth 2.1 or scoped API key; OpenAPI 3.1 spec, MCP server card and agent skill files; default sandbox provider for Kortix CompanyOS |
| Core | Core | CNY only, billed per second while running: vCPU ¥0.00003/vCPU-s (¥0.108/vCPU-h)CNY only, billed per second while running: vCPU ¥0.00003/vCPU-s (¥0.108/vCPU-h), memory ¥0.000015/GiB-s (¥0.054/GiB-h); 20 GB local disk free; templates and snapshots stored free up to 60 GB per account, then ¥0.0005/GB-h. 2 vCPU/4 GiB = ¥0.432/h. No USD price on the site, so the USD hourly figure is left as ? | No compute charge while paused; templates and snapshots free up to 60 GBNo compute charge while paused; templates and snapshots free up to 60 GB, then ¥0.0005/GB-h | States Running, Paused, Killed; timeout is the maximum lifetime counted from creationStates Running, Paused, Killed; timeout is the maximum lifetime counted from creation (default 5 minutes, changeable with set_timeout), after which the sandbox is stopped or, if configured, paused; an idle timeout fires when no client is connected and kills or pauses it; Paused keeps filesystem and memory with compute billing stopped and resumes in about 1 s, optionally automatically | VM | Jul 2025 | Agent sandbox from PPIO (Chinese AI cloud that also sells GPUs and model APIs)Agent sandbox from PPIO (Chinese AI cloud that also sells GPUs and model APIs); E2B-compatible API ("first in China") plus its own ppio-sandbox Python/JS SDK and CLI; start under 200 ms claimed; templates from a Docker image or Dockerfile; snapshots; secrets; code, browser and computer-use (virtual desktop) scenarios; regions cn-shanghai-1 (default) and cn-beijing-1; 1-8 vCPU and up to 8 GiB at CPU:memory 1:0.5 to 1:4; 50 concurrent sandboxes per user by default; launch post names Camel AI, OpenManus and Dify integrations |
| Core | Core | $0.09 | Disk only, $0.0002/GiB-hour; checkpoint storage billed separately | Default lifetime timeout 60 min (negative value = no deadline)Default lifetime timeout 60 min (negative value = no deadline); optional idle timeout up to 24 h, off by default; pause stops compute and keeps memory and processes, billed for disk only; paused time counts toward the timeout; checkpoints save the filesystem | VM | Sep 2026 | Up to 16 vCPU, 64 GiB RAM, 128 GiB disk per sandbox; boots an OCI/Docker image as root filesystemUp to 16 vCPU, 64 GiB RAM, 128 GiB disk per sandbox; boots an OCI/Docker image as root filesystem; thousands start concurrently in seconds; claims tens of thousands concurrent; SSH; Docker Compose; pause/resume with memory; checkpoints and fork; egress allowlists by domain/CIDR and private networks; GPUs (up to 8) coming soon; Python SDK and CLI; aimed at RL rollouts (Verifiers, prime-rl) |
| Core | Core | $0.13 | 20 GB paused storage included, then $0.05/GB-month | Pause keeps memory; saved memory and disk count as paused storage (retention not stated) | — | — | Free credit limits sandboxes to 2 vCPU / 4 GB for 60 minutes |
| Core | Core | $0.1656 | $0 (paused time not billed) | Opt-in auto-pause on idle, wakes on the next request; paused time not billedOpt-in auto-pause on idle, wakes on the next request; paused time not billed; paused sandbox kept 30 days; episode timeout 15 min by default; Free sessions up to 1 h; a sandbox that hits its spend cap is paused with files kept | — | Oct 2026 | 8 ms p50 ready claim (2 vCPU/4 GiB, n=100); 104 ms wake from pause; snapshot copied to 1,000 sandboxes in ~1 s8 ms p50 ready claim (2 vCPU/4 GiB, n=100); 104 ms wake from pause; snapshot copied to 1,000 sandboxes in ~1 s; content-hashed snapshots for reproducible evals; validity report per run; quotes and hard spend caps; outbound allow-list; secrets bound to hosts; GitHub Actions and GitLab CI runners; CLI with MCP server, TypeScript and Python SDKs, REST; Inspect, Harbor, SWE-bench adapters; US East only; own-cloud deployment on request |
| Core | Core | Priced per invocation, billed on the Logfire plan: Free tier 50,000 invocations/month onPriced per invocation, billed on the Logfire plan: Free tier 50,000 invocations/month on every Logfire plan; Pro $49/mo (add-on to Personal or Team) + $0.10 per 1,000 invocations after 50,000; included with Growth ($249/mo) and Enterprise Cloud; Advanced custom; $1k beta credit offered | $0 (billed per invocation; paused sessions stored as snapshots within plan$0 (billed per invocation; paused sessions stored as snapshots within plan limits) | Nothing to boot or keep warm; a session can be paused at a tool callNothing to boot or keep warm; a session can be paused at a tool call, snapshotted (interpreter state, e.g. 4.2 KB) and resumed later; snapshots kept 7 days on Free and Pro (2 GiB / 128 snapshots per project) | Container | Nov 2026 | Cold start in milliseconds; Python only (Monty subset, no third-party packages)Cold start in milliseconds; Python only (Monty subset, no third-party packages); code can call only the host functions the app exposes, no files, network or env vars unless granted; limits 50 ms CPU per invocation, 20 MiB memory per session, 1,000 host calls per session, 300 session starts per 5 min; every run traced in Logfire; Monty open source (MIT, 8,926 GitHub stars); self-hosted version planned for Logfire Enterprise |
| Core | Core | $0.25 | $0 compute while paused; volume storage price ? | Pause releases compute and keeps the sandbox's shared volumePause releases compute and keeps the sandbox's shared volume; a command sent to a paused sandbox resumes it; idle sandboxes are terminated after the template's idle timeout and every sandbox ends at the template's hard lifetime; persistence modes EPHEMERAL, DIRECTORY (live folder on object storage), SNAPSHOT (workspace at pause) or BOTH | gVisor | — | Part of the Quave ONE PaaS (formerly zCloud, then Quave Cloud)Part of the Quave ONE PaaS (formerly zCloud, then Quave Cloud); any Linux container image from immutable versioned templates; REST API and MCP tools (create, execute, pause, resume, terminate, shared volumes); idempotent create; egress allowlists by domain or IP range, internal networks blocked; per-command secrets; data on Quave object storage or your own S3-compatible bucket with your own key; starts in seconds when capacity is warm; regions incl. US-5, BR-1, EU-3; BYOC on AWS, GCP, Azure, OCI or on-premise via Connect |
| Core | Core | $0.054 | $0 compute; disk $0.20/GiB-month | Auto-suspend after 30 idle min (15 on Free; paid plans can turn it off)Auto-suspend after 30 idle min (15 on Free; paid plans can turn it off); suspend keeps memory and processes when the provider can, else stops and keeps the disk; next command wakes it; background jobs up to 24 h | VM | Nov 2026 | Debian 13 with Claude Code, Codex, Gemini CLI, OpenCode, Cursor CLI and Copilot CLI preinstalledDebian 13 with Claude Code, Codex, Gemini CLI, OpenCode, Cursor CLI and Copilot CLI preinstalled; persistent /data disk; snapshots and fork (disk only); HTTPS URL per published port; CLI (npm quivvo), REST API with OpenAPI, hosted MCP server; regions ams (Amsterdam) and iad (Ashburn); max 8 vCPU/16 GiB; sign-up queue with invite links |
| Core | Core | $0.4167 | No paused state: an idle sandbox keeps billing for memory in use until the idleNo paused state: an idle sandbox keeps billing for memory in use until the idle timeout destroys it; checkpoint storage rate not stated | No pause; destroyed automatically after the idle timeout (Hobby/Pro default 30 minNo pause; destroyed automatically after the idle timeout (Hobby/Pro default 30 min, 1-120 min or disabled with 0; Trial/Free default 5 min, max 5); commands, heartbeats, SSH and port forwarding reset the timer; checkpoints save the disk for later boots or forks; billable until destroyed | VM | Jun 2026 | Ubuntu image with git, Node and common coding agents preinstalled; TypeScript SDK, CLI and dashboardUbuntu image with git, Node and common coding agents preinstalled; TypeScript SDK, CLI and dashboard; exec with detach/reattach, stdin, files API; checkpoints and forks of running sandboxes; SSH and port forwarding from the CLI; Docker inside sandboxes; private networking (PRIVATE mode) or ISOLATED by default; public domains; regions incl. us-west2 default and us-east4; sizes up to 2 vCPU/2 GB (Trial/Free), 8/8 (Hobby), 32/32 (Pro); 10/10/50/100 sandboxes per environment |
| Core | Core | CNY only, per second, settled hourly: vCPU ¥0.03/vCPU-hourCNY only, per second, settled hourly: vCPU ¥0.03/vCPU-hour, memory ¥0.01/GiB-hour (rounded up to whole GiB), storage ¥0.50/GiB-month beyond 32 GiB free per sandbox (on the declared workspace size), egress ¥0.60/GB; 2 vCPU / 4 GiB = ¥0.10/h (the site's own example); no monthly fee or minimum; ¥50 sign-up balance; WeChat Pay top-ups; model tokens are not resold. No USD price on the site (its comparison table uses $1 ≈ ¥6.82, about $0.0147/h), so the USD hourly figure is left as ? | ¥0 compute while paused or stopped; storage ¥0.50/GiB-month beyond 32 GiB free¥0 compute while paused or stopped; storage ¥0.50/GiB-month beyond 32 GiB free per sandbox | States running, paused, stopped, destroyed: pause keeps processes and memory in placeStates running, paused, stopped, destroyed: pause keeps processes and memory in place (resume measured at 7 ms); stop keeps the workspace, the 4 GiB overlay disk and the IP and can restart in place; optional timeout auto-stop; when the balance runs out, sandboxes are cold-stopped, then destroyed after a 3-day grace period (1 day for self-serve and trial orgs) | VM | — | Sandbox cloud for agent workloads; 8 official agent images including Claude Code (default), Codex, OpenCodeSandbox cloud for agent workloads; 8 official agent images including Claude Code (default), Codex, OpenCode, Cursor CLI, Kimi Code and DeepSeek Harness; model API keys injected by the platform, kept out of images and logs; platform-recorded audit event stream (lifecycle, egress, API calls) queryable by sandbox, event type, trust level and risk level; terminal, files, SSH and Chrome Remote Desktop; E2B-compatible management API (create, list, delete tested; commands and files not yet); Harbor eval integration; CLI for Linux x86_64; default 2 vCPU / 2 GiB; Chinese-language site with a Beijing ICP filing; GitHub org RayrunOS created 2026-11-01 |
| Core | Core | $0.316 | $0 compute when suspended; snapshot storage $0.08/GB-month | Auto-suspend after 15 min idle by default (up to 24 h, or off)Auto-suspend after 15 min idle by default (up to 24 h, or off); a suspended sandbox keeps its filesystem, auto-resumes on the next command and stays resumable for 30 days; runtime timeout 24 h by default (max 30 days, or disabled); runtime snapshots keep running processes; termination deletes the filesystem | VM | Nov 2026 | Starts in under a second claim; fork from snapshot with running processes; Docker and Docker Compose insideStarts in under a second claim; fork from snapshot with running processes; Docker and Docker Compose inside; custom images; TypeScript and Python SDKs, CLI; outbound allowlist; credential brokering and scoped API keys in early access; 1,000 concurrent by default on Pro/Scale; Oregon only; no public URLs or private networking yet (beta); up to 8 CPU/16 GB |
| Core | Core | Per request plans: Hobby free (100,000 requests/mo, 100 GB transfer, 30 s limitPer request plans: Hobby free (100,000 requests/mo, 100 GB transfer, 30 s limit, 3 custom runtimes); Pro $250/mo (1,000,000 requests, 500 GB transfer, up to 15 min, 50 custom runtimes; usage-based pricing above, not published); Enterprise custom, including self-hosting | $0 (no idle sandboxes; billed per request plan) | Stateless per request: each call executes code and returnsStateless per request: each call executes code and returns; max execution time 30 s (Hobby) or up to 15 min (Pro); no persistent sandbox | WebAssembly | Apr 2024 | Runs untrusted Python, JavaScript, TypeScript, Ruby and PHP; code starts in under 10 ms, no cold startsRuns untrusted Python, JavaScript, TypeScript, Ruby and PHP; code starts in under 10 ms, no cold starts; Python, TypeScript and Go SDKs and REST API; custom runtimes with third-party packages (Python and JS); per-request env vars, files and HTTP-only network access, secrets and HTTP credential injection; /mnt/req read-write, rest read-only; FAQ says memory limited to 128 MiB and 30 s per invocation (conflicts with Pro's 15 min); SOC 2 Type 1 (Dec 2024); self-host on Enterprise |
| Core | Core | $0.3168 | $0 CPU/RAM; storage $0.00034236/GB-hr | Suspend stops CPU and RAM billing; storage keeps billing | VM | — | Free trial |
| Core | Core | $0.1656 | Paused: memory + disk ($0.0162/GiB-h + $0.000108/GiB-h), no vCPUPaused: memory + disk ($0.0162/GiB-h + $0.000108/GiB-h), no vCPU; suspended or shut down: disk only ($0.000108/GiB-h); deleted: $0 | Auto-suspend after a configured idle timeout with no inbound HTTP traffic (idle scanAuto-suspend after a configured idle timeout with no inbound HTTP traffic (idle scan every 60 s) and wake on the next request, which is proxied after restore; Paused keeps memory and disk; Suspended or shut down keeps disk only; checkpoints capture the filesystem and running processes; memory auto-scaling | — | Jul 2026 | "Execution layer for AI agents": runtimes from the CLI, Python and TypeScript SDKs, REST API or hosted MCP server"Execution layer for AI agents": runtimes from the CLI, Python and TypeScript SDKs, REST API or hosted MCP server; prompt compression of tool output to save tokens; secret stubs with real keys added at the egress gateway; egress allow rules; published HTTPS ports; SSH access (Aug 2026); checkpoints; Cloud Agents and Exo Harness; open-source Errand agent; Trust Center lists SOC 2 Type II and ISO 27001 as in progress |
| Core | Core | $0.08 | $0 compute when paused; paused storage $0.08/GB-month (memory and disk) | Auto-pause after 300 s with nothing happening (idlePauseSeconds 60-86,400 or 0 forAuto-pause after 300 s with nothing happening (idlePauseSeconds 60-86,400 or 0 for never), keeping memory and processes; wakes on the next command, file call or preview visit; stopped sandboxes keep their disk until deleted; paused/stopped disks kept 30 days after credit runs out (7 days for accounts without a purchase or card) | VM | Sep 2026 | Pause keeps memory; forks of 1-10 running copies with memory and processes; custom images from DockerfilePause keeps memory; forks of 1-10 running copies with memory and processes; custom images from Dockerfile; volumes backed up daily; SSH; desktop; code interpreter; Docker; preview URLs, custom domains, TCP ports, WireGuard private networks; secrets added at egress; bucket mounts; MCP server catalog; drop-in imports for E2B, Daytona, Vercel Sandbox, Blaxel, Modal, Morph and Cloudflare code; SDKs for JavaScript, Python, Go, Ruby, Java; up to 16 vCPU / 64 GiB per sandbox; agents can open an account and pay in USDC; 99% API uptime SLA; Runtime LLC, Denver |
| Core | Core | No public rates; prepaid credits topped up through StripeNo public rates; prepaid credits topped up through Stripe, usage metered as sandbox runtime seconds, model tokens, storage byte-seconds and tool executions; the X profile (not the site) says $20 free credits on sign-up | Pause avoids idle compute per docs; storage rate not published | Pause snapshots the sandbox and resume brings back the same processes and memoryPause snapshots the sandbox and resume brings back the same processes and memory; a command or dev-URL request wakes a paused sandbox; idleTimeout set per sandbox at create (docs example 1800 s); sandbox-local files are removed on destroy, mounted workspaces persist; named checkpoints, restore and clone are not in the public API | VM | May 2026 | Agent platform: sandboxes (base-ubuntu and desktop-ubuntu templates for browser/GUI automation)Agent platform: sandboxes (base-ubuntu and desktop-ubuntu templates for browser/GUI automation), persistent org Workspaces mounted at /workspace, agents with schedules, webhook triggers and Telegram channels, Tool Hub, RunMesh private networking to local devices, local-mac execution mode; SSH, protected dev URLs per port, VNC readiness; resources set per sandbox (docs example 2 vCPU / 4GB / 10GB disk); warm-pool claims or cold boot; TypeScript SDK @runtools-ai/sdk, CLI and REST API; the runtools.ai homepage now sells custom software and AI-systems consulting and only mentions 'isolated compute, persistent workspaces' as its production infrastructure |
| Core | Core | $0.062 | $0 while sleeping or paused (CPU, memory and disk not billed)$0 while sleeping or paused (CPU, memory and disk not billed); volumes $0.000411/GiB-hr | Sleep (automatic after an idle period, wakes on trafficSleep (automatic after an idle period, wakes on traffic, commands or file ops) or pause (explicit resume) both save memory, processes and filesystem; no runtime limit; automatic checkpoints to S3 every few minutes; optional max lifetime ends in termination; when credits run out Sailboxes are paused, not terminated | VM | Apr 2026 | Startup <3 s; up to 64 vCPU, 256 GiB RAM, 1 TB disk per Sailbox; memory grows and shrinks with useStartup <3 s; up to 64 vCPU, 256 GiB RAM, 1 TB disk per Sailbox; memory grows and shrinks with use; live migration keeps processes and connections; checkpoints of memory+disk and copies from checkpoints; SSH, CLI, Python and TypeScript SDKs; egress policies, credential injection and vaults; custom domains; Docker and systemd; SOC 2 Type II, HIPAA BAA on Enterprise; Sail also sells model inference |
| Core | Core | Individual $20/month for up to three Computers and 20 GB pooled storageIndividual $20/month for up to three Computers and 20 GB pooled storage; Enterprise from $300/month | — | Keeps files, repos, packages, browser sessions and running services between sessions | — | Jul 2026 | Persistent cloud computers for agents; OpenCode Computer and Hermes Computer environmentsPersistent cloud computers for agents; OpenCode Computer and Hermes Computer environments; browser use, MCP, secrets vault, network policy, web terminal with root; SSO and VPC deployment on Enterprise |
| Core | Core | $0.09 | $0 (no paused state; sandbox destroyed at timeout)$0 (no paused state; sandbox destroyed at timeout); snapshot storage price not stated | No pause; sandbox deleted at timeout_minutes (default 15, max 1440 min after creationNo pause; sandbox deleted at timeout_minutes (default 15, max 1440 min after creation, extensions included); snapshots of filesystem kept 7 days (max 3 per account), restore into a new sandbox without processes | VM | Aug 2026 | REST API plus MCP server; Python and JS/TS SDKs; warm pool of 1 machine gives ~1 s startREST API plus MCP server; Python and JS/TS SDKs; warm pool of 1 machine gives ~1 s start, cold create ~24-47 s (measured, published); 20 concurrent sandboxes; preview URLs; Python 3.12, Node 22 preinstalled; runs unprivileged (no root); operated by productivity-boost.com Betriebs UG & Co. KG, Passau, Germany |
| Core | Core | $0.06 | $0 compute; retained storage $0.02/GiB-month | Pause checkpoints the writable RootFS and releases compute (fresh processes on resumePause checkpoints the writable RootFS and releases compute (fresh processes on resume; experimental memory checkpoints); auto_resume on inbound access; paused sandboxes kept indefinitely; a running idle sandbox still bills | gVisor | Feb 2026 | Persistent, auto-growing RootFS (up to 64 TiB sparse); snapshots and forks; SSH gateway; Docker in sandboxPersistent, auto-growing RootFS (up to 64 TiB sparse); snapshots and forks; SSH gateway; Docker in sandbox; files readable while paused; egress policy and credential injection; s0 CLI, Python SDK; 20 active sandboxes per team; self-hosting no longer supported (old Apache-2.0 repo archived, 87 stars), BYOC planned |
| Core | Core | Monthly plans sold through RapidAPI with hard quotas and no overage: Basic $0 (500Monthly plans sold through RapidAPI with hard quotas and no overage: Basic $0 (500 executions/mo, 30 s max CPU timeout); Pro $19/mo (10,000 executions, 60 s, 5 concurrent sessions, pip/npm install, async); Ultra $49/mo (50,000 executions, 300 s, 20 sessions). Billed per execution, not by time; sandbox CPU/RAM size not published | $0 (sessions expire after at most 30 idle minutes) | Plain executions keep no data; async results kept 24 hPlain executions keep no data; async results kept 24 h; sessions keep files and installed packages until an idle TTL (max 30 min), renewed on each execute or install; every call starts a fresh interpreter | gVisor | Apr 2026 | 8 runtimes: Python 3.12, Node 22, TypeScript 5.4, Go 1.22, Java 21, C and C++ (GCC 14), Bash 5.28 runtimes: Python 3.12, Node 22, TypeScript 5.4, Go 1.22, Java 21, C and C++ (GCC 14), Bash 5.2; sync, SSE streaming, async polling and session modes; batch up to 10/50/100; pip and npm through a cached mirror of the top 1k packages; pre-warmed pools with p50 cold start under 100 ms (changelog 1.2.0); MCP server at mcp.sandboxapi.dev for Claude Code, Codex, Cursor, VS Code, Antigravity; self-hosted edition being scoped for late 2026 |
| Core | Core | $0.018 | $0 (stopped sandboxes kept free for 14 days$0 (stopped sandboxes kept free for 14 days; counts toward plan's included storage) | Stop snapshots memory, device state and diskStop snapshots memory, device state and disk, then resume restores the same processes (falls back to a clean boot from disk if the memory image is missing, the size changes, or no matching host frees up in 2 min); auto-pause when idle (after 30 min on the $20 plan) and wake on the next request; auto-stop deadline default 1 h, or never; disk snapshots every 10 min while running; stopped sandboxes kept free for 14 days after they last stopped | VM | Oct 2026 | Ubuntu 24.04 VMs with root, systemd, Docker, nested KVM; pause with memoryUbuntu 24.04 VMs with root, systemd, Docker, nested KVM; pause with memory; instant copy-on-write forks (reflinked disks); named snapshots as templates, shareable by code; HTTPS URL for any port; SSH/SCP over HTTPS plus IPv4 port and public IPv6; XFCE + Chrome desktop over VNC for computer use; Claude Code, Codex, OpenCode preinstalled and driven by API; signed webhooks and event stream; scoped API keys; Daytona-compatible API; regions US East (Ashburn) and EU Central (Helsinki); no GPUs, no SLA during early access; REST API with OpenAPI, sbx CLI, TypeScript and Python SDKs, MCP server; North Bay Opco I, Inc. (Delaware, Miami address) |
| Core | Core | Flat monthly plans, no free plan: Plus $5/mo (5 sandboxes, 2 GB storageFlat monthly plans, no free plan: Plus $5/mo (5 sandboxes, 2 GB storage, 4 h session timeout, up to 4 vCPU); Pro $15 to $100/mo in four tiers (Starter $15: 10 sandboxes, 2 GB; Growth $30: 25, 10 GB; Scale $60: 60, 30 GB; Business $100: 150, 100 GB; 8 h session timeout, up to 8 vCPU); yearly saves 20%; Enterprise custom. The billing docs page still describes an older credit scheme (1 credit = 1 vCPU-hour or 1 GiB-hour; Plus $29/mo with 400 credits, $0.10/credit overage) | Flat monthly plans | Auto-stop after inactivity: default 1,800 s (30 minutes), set with activeDeadlineSecondsAuto-stop after inactivity: default 1,800 s (30 minutes), set with activeDeadlineSeconds, timer resets on each command; states Creating, Running, Stopped, Deleted; a stopped sandbox is kept until deleted; persistent SSD volumes survive; plan session timeout 4 h (Plus) or 8 h (Pro) | Container | Jan 2025 | Docker sandboxes with SSD-backed, POSIX-compatible persistent volumes for Claude Code, Codex and MCP serversDocker sandboxes with SSD-backed, POSIX-compatible persistent volumes for Claude Code, Codex and MCP servers; JavaScript/TypeScript and Python SDKs, REST and OpenAPI, CLI login; terminal, browser automation, port proxy; default 0.5 vCPU and 1 GB per sandbox, maximum 4 vCPU and 8 GB; any Docker image (default ubuntu:22.04); GitHub sandock-ai/sandock (SDKs, MIT, 4 stars) |
| Core | Core | No public price; the README warns that quick starts create "billable or quota-boundNo public price; the README warns that quick starts create "billable or quota-bound resources" | — | E2B-style lifecycle timeout in seconds (set_timeout, refresh, heartbeat)E2B-style lifecycle timeout in seconds (set_timeout, refresh, heartbeat); pause, resume and connect; snapshot APIs not exposed because "the underlying platform does not support them" | Container | Apr 2026 | E2B-style SDKs for Python, Node and Go (MIT, 0 stars each) plus seacloud-cliE2B-style SDKs for Python, Node and Go (MIT, 0 stars each) plus seacloud-cli; official templates base, code-interpreter (Python, JavaScript, TypeScript, Bash, R, Java), claude and codex (CLIs preinstalled); custom tpl- templates built from image steps; files, commands, PTY, git, public proxy URL per port; volumes on NFS, block or object storage; production gateway hosted under seaart.ai (SeaArt, an AI art platform); no public website |
| Core | Core | Compute and AI metered as dollar credits (1 credit = $1) under an account-wide hardCompute and AI metered as dollar credits (1 credit = $1) under an account-wide hard budget cap: Free $0 ($2 trial credits, 1 Small server, 5 GB, 10 GB egress, hard stop); Starter $29/mo ($10 credits, 2 Small servers, 5 GB per server, 100 GB egress); Pro $99/mo ($40 credits, 3 Small or Medium servers, 20 GB, 250 GB egress); Business $299/mo ($150 credits, 10 servers up to Large, 100 GB, 1 TB egress); Enterprise custom with BYOC; overage about 1.2-1.4x the base rate; egress overage $0.12, $0.10 or $0.08 per GB; annual billing about 2 months free. No per-hour compute rate and no CPU or memory per tier published (tiers differ by project count and storage) | $0 compute while asleep (idle servers scale to zero)$0 compute while asleep (idle servers scale to zero); storage is included in the plan | Servers are persistent: each keeps its workspace and state between tasks on a durableServers are persistent: each keeps its workspace and state between tasks on a durable volume (/home/node/workspace), and idle servers scale to zero (status running, sleeping or building); the account holder can pause projects, kill processes and delete servers | — | Jun 2026 | MCP server (34 tools covering servers, projects, files, shell commands, secrets, buildsMCP server (34 tools covering servers, projects, files, shell commands, secrets, builds, deploys and logs) and REST API for ChatGPT, Claude, Cursor, Codex CLI, GitHub Copilot and Gemini CLI; agents provision servers with create_server and no human approval; an on-server build agent works from a goal; projects get <slug>.apps.server4agent.com URLs, private by default or team-only; browser VS Code with git history; encrypted secrets vault; WAF and DDoS protection; regions us-east, us-west, eu-west; one unauthenticated call publishes a static page; the site calls itself 'not a sandbox primitive'; no SOC 2 report yet; Server4Agent, Inc., Sunnyvale |
| Core | Core | $0.14 | $0 compute while parked or suspended$0 compute while parked or suspended; retained storage inside the plan allowance, then $0.15/GiB-month | A running workspace parks seconds after its last call (memory compressed in placeA running workspace parks seconds after its last call (memory compressed in place; it counts 256 MiB for about 15 s, then nothing) and the next call wakes it; suspend and resume keep memory and running processes; suspended workspaces are kept as long as you want within the retained-storage allowance; fork with memory and processes; opt-in idle retention removes unused workspaces | VM | — | Persistent workspaces addressed by your own key (e.g. customer-42/main)Persistent workspaces addressed by your own key (e.g. customer-42/main), created on the first call from versioned templates; TypeScript and Python SDKs, the shard CLI, MCP server and HTTP API; E2B-compatible API; tools for exec, files, terminals, git, browser and computer use with a desktop; private HTTPS URL per port that wakes the workspace; GPU through Shardflux's own Vulkan driver with CUDA Driver API and NVENC/NVDEC on Startup and Scale; region AWS us-west-2 (Oregon) only; by Helios One Oy |
| Core | Core | Compute $0.12 per compute-hour on the standard 8 vCPU / 4 GiB shapeCompute $0.12 per compute-hour on the standard 8 vCPU / 4 GiB shape, billed per second only while a command runs (active only); services $0.03/h (2 vCPU / 2 GiB) or $0.06/h (4 vCPU / 4 GiB) while awake; storage $0.08/GiB-month (first 5 GiB free per Space); egress $0.03/GB (first 10 GiB free per Space); plans Shark Pro $5/month (includes $5 usage), Studio $25/month (includes $25), Scale $100/month (includes $100); no 2 vCPU / 4 GiB compute shape | $0 compute; storage $0.08/GiB-month after 5 GiB free per Space | No VM between commands: each command runs in a fresh microVM that is destroyed when itNo VM between commands: each command runs in a fresh microVM that is destroyed when it ends; the Space (files, databases) persists; services sleep about 10 s after their last request and wake on the next (Space wakes in 58 ms); hibernate and wake; sync exec timeout 120 s default, 900 s max (7200 s streamed) | VM | — | 44 ms p50 command in a fresh microVM, 65 ms create to first command44 ms p50 command in a fresh microVM, 65 ms create to first command, 342 ms for a burst of 24 (own benchmark on a Hetzner AX41); fork a Space (up to 50 at once); per-Space Postgres 16/17 or Redis 7 that sleep and wake; secrets swapped in at the egress proxy, never enter the VM; preview URLs and long-running processes in beta; Claude Code through a shell shim, Codex, OpenCode (shark code); Python and TypeScript SDKs; sharkfn grep/glob over MCP; node hel1 (Helsinki); no GPUs; process memory and shell state don't persist between commands; a product of SharkWorks, Mexico |
| Core | Core | $0.02 | $0.50/month per x1 slot when parked | Default stop mode: pauses 120 s after the last SSH session disconnectsDefault stop mode: pauses 120 s after the last SSH session disconnects, with a memory snapshot (processes, shell history, TCP state), and resumes on connect in ~3 s; keepalive (always on), wakeup (starts on an HTTPS request, stops after 5 min without TCP activity) and cron modes; boxes auto-stop below $5 balance and are deleted at $0 | VM | Jan 2026 | SSH only, no signup (SSH key fingerprint is the account); boot from prepared OCI imagesSSH only, no signup (SSH key fingerprint is the account); boot from prepared OCI images; duplicate a box including its memory snapshot; HTTPS URL with automatic TLS, email endpoint, custom domains and public IPv6 per box; works with VS Code Remote, Zed, mosh, scp, sftp; Hetzner bare metal in Finland, one region; payments via Paddle |
| Core | Core | Self-hosted license: Individual $25/month (2 daemons, unlimited VMs, 14-day free trialSelf-hosted license: Individual $25/month (2 daemons, unlimited VMs, 14-day free trial, not for shared infrastructure); Team $125/month for 5 seats + $25/month per extra seat; Platform $250/month per server for shared infrastructure and SaaS; custom plans | N/a (your own hardware; no metering) | Self-hosted: Sandboxes are ephemeral microVMs launched and torn down by APISelf-hosted: Sandboxes are ephemeral microVMs launched and torn down by API; Services are long-lived systemd microVMs; commit a VM's disk and fork or clone it; no metering | VM | Sep 2025 | Cold boot under 1 s with ZFS snapshots; kernel + systemd in under 300 ms with optimised images (blog 2026-01-14)Cold boot under 1 s with ZFS snapshots; kernel + systemd in under 300 ms with optimised images (blog 2026-01-14); REST API, Go SDK, TypeScript SDK, CLI; default-deny egress proxy with Bearer/Basic/OAuth secret injection; Slicer for Mac (preview 2026-02-11); GPU/PCI passthrough via VFIO; Ubuntu LTS and Rocky 9 images; K3s clusters; bare metal, mini PCs, Arm, nested virtualization on GCE, Azure, OCI, DigitalOcean and AWS C8i/M8i/R8i; used internally at OpenFaaS since 2022 |
| Core | Core | $0.171 | Site says paused machines stop counting against the plan (paused computeSite says paused machines stop counting against the plan (paused compute billing not stated explicitly); snapshot storage $0.05/GB-month beyond 10 GB free | idleTimeoutMs (15 min suggested) then lifecycle onTimeout 'pause' (defaultidleTimeoutMs (15 min suggested) then lifecycle onTimeout 'pause' (default; saves state, resumes on connect) or 'kill'; optional park (CPUs frozen, slot held) escalating to full pause; stop keeps files but not processes, start is a cold boot; max session 1 h Free, 5 h Starter, 24 h Professional, unlimited Team/Enterprise | VM | Aug 2026 | Ready sandbox in under 90 ms (claim); snapshot, fork and rewind running sandboxesReady sandbox in under 90 ms (claim); snapshot, fork and rewind running sandboxes; bring your own image, persistent filesystem and volumes; scale 1-16 vCPU without losing state; SDKs TypeScript, Python, Go, Rust, C++ over one REST + WebSocket protocol; also sells stealth Chrome browsers, Linux desktop VMs, proxies and captcha solving; migration guides from E2B, Daytona, Browserbase; concurrency 1/2/10/50+ sandboxes by plan |
| Core | Core | No published rates; private beta with $20 free credit and no card requiredNo published rates; private beta with $20 free credit and no card required; sandbox responses show billing_model 'payg' | Paused sandbox keeps its disk allocation; no rates published | Auto-pause after idle_timeout_s (default 300 s; runs and commands count as activityAuto-pause after idle_timeout_s (default 300 s; runs and commands count as activity; a sandbox with a terminal attached never idles out); pause snapshots memory, device state and disk and frees CPU and memory, but running processes and interpreter state start fresh after resume; destroyed at max_lifetime_s (set at create, capped by account policy) and after 7 days paused | VM | — | HTTP API (create, run Python/Node/Bash, exec, pause, resume, destroy)HTTP API (create, run Python/Node/Bash, exec, pause, resume, destroy); live WebSocket terminals, VS Code in the browser (1-hour IDE sessions) and signed port-preview links (1 s to 24 h); sizes nano 1 vCPU/256 MiB/1 GiB, small 2 vCPU/512 MiB/10 GiB, medium 2 vCPU/1 GiB/20 GiB, large 2 vCPU/4 GiB/20 GiB; images Alpine 3.22, Debian 13, Ubuntu 24.04, plus a Debian 'agents' template with Claude Code, Codex, OpenCode, OpenClaw and Cursor CLI (large only); spawn CLI uploads a project and runs Claude Code in a sandbox (spawn claude / spawn pull); internet access can be switched off; 100 sandboxes per account and 30 s per run by default; official Python/TypeScript SDKs, browser automation and custom images not available yet |
| Core | Core | $0.1492 | Site says compute stops only if the runtime actually suspendsSite says compute stops only if the runtime actually suspends; pause availability shown as Unknown; retained storage charges not listed | Each sandbox has a configured lifetime (e.gEach sandbox has a configured lifetime (e.g. timeout_seconds) and can be stopped by dashboard or API; max runtime 45 min Hobby / 24 h Pro and Team; durable sandboxes (Pro 20, lifetime 3 days; Team 100, 7 days); workspace snapshots keep files but not memory; pause/resume and memory preservation listed as depending on runtime (availability shown as 'Unknown'); shared sandboxes run on Spot and can be preempted | gVisor | Jan 2026 | Pre-warmed pods (homepage); code and shell execution, files and artifacts; browser-ready environmentsPre-warmed pods (homepage); code and shell execution, files and artifacts; browser-ready environments; TypeScript and Python SDKs (@stateset/sandbox-sdk 2.13.0) and REST API; run budgets, scoped API keys, team roles, encrypted secrets, audit logs; commerce and operations sample workflows; Team add-ons: private networking, VPC peering, SSO/SAML; region selection not confirmed; part of StateSet |
| Core | Core | $0.0812 | $0 compute when stopped or paused; snapshots count toward persistent storage$0 compute when stopped or paused; snapshots count toward persistent storage ($0.00002727/GiB-hr above 10 GiB free) | Runs 5 min by default (timeout settable and resettable)Runs 5 min by default (timeout settable and resettable); maximum runtime 1 h by default (more on request); pause (public beta) saves filesystem and memory including running processes, connect() resumes; auto-pause after 10 min inactivity (beta); kill deletes | VM | — | Startup under 200 ms; fully compatible with the E2B SDK (docs use @e2b/code-interpreter); pause/resume with memoryStartup under 200 ms; fully compatible with the E2B SDK (docs use @e2b/code-interpreter); pause/resume with memory; custom templates; code interpreter; Git integration, secret injection, resource mounting, object-storage buckets; guides for Claude Code, Codex, OpenCode and Amp; MCP; part of SUFY's cloud (CDN, object storage, model API) |
| Core | Core | $0.3168 | $0 compute; disk $0.15/GB-month | Stops after 30 s idle (an open outbound connection counts as use for up to 10 min)Stops after 30 s idle (an open outbound connection counts as use for up to 10 min), becomes warm and later cold (disk fetched back on wake); any exec, file, preview or schedule request wakes it; processes end on stop, disk and checkpoints kept; hold keeps it stopped | VM | Feb 2026 | Default sandbox 4 vCPU arm64 / 8 GiB / 20 GiB disk, Debian 13 as root (homepage still says 2 vCPU / 2 GiB)Default sandbox 4 vCPU arm64 / 8 GiB / 20 GiB disk, Debian 13 as root (homepage still says 2 vCPU / 2 GiB); ready in under a second; HTTP API and WebSocket shell; secrets injected outside the machine; preview URLs; Durables (persistent agents built on Pi, any model with your key); Apps deployed from GitHub; bare metal in Sydney, Australia only; 25 running and 100 kept per workspace; shuru runtime Apache-2.0, 894 GitHub stars |
| Core | Core | $0.1656 | $0 compute while paused; storage $0.000108/GiB-hr for retained diffs per$0 compute while paused; storage $0.000108/GiB-hr for retained diffs per pricing page (live billing API marks storage non-billable) | Pause checkpoints the whole machine (files, packagesPause checkpoints the whole machine (files, packages, memory snapshot) and stops compute billing; auto-pause after timeoutSeconds of active time or idleTimeoutSeconds with no activity; next command, file operation or connect resumes it; no fixed session limit on Developer and Pro; fork from snapshots | VM | Jul 2026 | Median 0.41 s create-to-first-command (ComputeSDK benchmark cited on site); NVMe disks and perpetual memory snapshotsMedian 0.41 s create-to-first-command (ComputeSDK benchmark cited on site); NVMe disks and perpetual memory snapshots; fork any machine; buckets; Docker and computer use; custom templates; egress rules, credential swap broker, request log; preview URLs; TypeScript and Python SDKs, CLI, MCP server and REST (Apache-2.0, github.com/superserve-ai/superserve, 512 stars); runs Claude Code, Codex, OpenCode and others; BYOC in AWS/GCP/Azure for early partners; HIPAA compliant, SOC 2 Type II in audit |
| Core | Core | Not published: the pricing section says the FreeNot published: the pricing section says the Free, Team and Enterprise (BYOC) plans are planned and prices, quotas and SLA are not final | — | Optional timeout auto-pauses an idle running sandbox (SIGSTOP / cgroup freezer: memoryOptional timeout auto-pauses an idle running sandbox (SIGSTOP / cgroup freezer: memory, file descriptors and sockets kept, resume in milliseconds); optional ttl destroys it regardless of state; stop keeps the filesystem but ends processes; destroyed is final | Container | May 2026 | Cold start under 50 ms claimed (warm pool; P95 about 30 ms with runc); streaming PTY with .cast session recordingsCold start under 50 ms claimed (warm pool; P95 about 30 ms with runc); streaming PTY with .cast session recordings; built-in Chromium with CDP; preview URLs with signing and TTL; network policy open, allowlist or blocked; overlayfs plus workspace volume persistence; SDKs TypeScript, Python, Go, .NET and the Go CLI tsb; self-host guides (Docker Compose, systemd) although the platform source repo is private; Chinese-language site by Talon Computing Inc.; the customer logos and testimonial (acme.dev, codex labs, raven ai) look like placeholders |
| Core | Core | $0.0491 | $0 compute when stopped; disk kept until deleted (disk rate $0.000734/GB-hour$0 compute when stopped; disk kept until deleted (disk rate $0.000734/GB-hour is listed for running sandboxes; stopped-disk charge not stated) | idleTimeoutSeconds stops after inactivity and maxLifetimeSeconds stops after a set timeidleTimeoutSeconds stops after inactivity and maxLifetimeSeconds stops after a set time (CLI defaults 900 s idle, 3600 s lifetime), disk kept; stopped sandbox bills no compute and keeps workspace files, agent conversations and snapshots but loses running processes; deleted after deleteAfterStoppedSeconds (default 30 days); snapshots and forks (branch up to 16) | Container | — | Coding agents installed: Claude Code, Codex, OpenCode, Hermes, OpenClaw, Kimi Code, Pi, Prime AgentCoding agents installed: Claude Code, Codex, OpenCode, Hermes, OpenClaw, Kimi Code, Pi, Prime Agent; box.prompt() runs an agent turn and returns cost; fleets with one budget and spend caps; TypeScript and Python SDKs, CLI, SSH; snapshots and branching; traces and evals; GPU leases; HIPAA placement and confidential hardware with attestation; migration guides from E2B, Daytona, Modal; SOC 2 Type II |
| Core | Core | CNY only, billed per second while running: CPU ¥0.000081/core-s (¥0.2916/core-h)CNY only, billed per second while running: CPU ¥0.000081/core-s (¥0.2916/core-h), memory ¥0.000025/GiB-s (¥0.09/GiB-h); system disk ¥0.0021/GiB-h beyond 15 GiB free while running; paused-state storage ¥0.0021/GiB-h; image pre-warming ¥0.00025/GiB-h storage + ¥0.00231/GiB-h use; COS/CFS billed by those products. 2 vCPU/4 GiB = ¥0.9432/h (Tencent's own example). No USD price found, so the USD hourly figure is left as ? | Paused: no CPU or memory charge; storage ¥0.0021/GiB-h on disk used (plus thePaused: no CPU or memory charge; storage ¥0.0021/GiB-h on disk used (plus the declared memory size when paused with memory), with no 15 GiB free allowance | Timeout set at creation (default 5 minutes) and changeable while runningTimeout set at creation (default 5 minutes) and changeable while running; the instance is stopped when it expires; pause and resume with or without memory; idle auto-pause and auto-resume on the next data request with keepalive (public beta since 2026-04-09); custom sandboxes have a persistent mode not reclaimed by timeout (2026-03-17); Mobile sandboxes run up to 15 days | — | Sep 2025 | Sandbox types Browser, Code Interpreter, SWE-bench, OSWorld (Ubuntu desktop), All-In-One, Mobile (ADB)Sandbox types Browser, Code Interpreter, SWE-bench, OSWorld (Ubuntu desktop), All-In-One, Mobile (ADB), Android World and Custom images; E2B-compatible API and E2B-format API keys; agr CLI; about 100 ms start and "hundreds of thousands of instances per minute" claimed; storage mounts COS, CFS, TurboCFS, JuiceFS, AgentBucket, Agent CBS; network policies; image pre-warming; process-level snapshots; custom sizes 0.5-32 vCPU with CPU:memory 1:1, 1:2 or 2:1; default quota per account and region 50 instances, 50 vCPU, 100 GiB, 20 paused; browser web terminal |
| Core | Core | $0.1656 | $0 compute while paused; disk $0.000108/GiB-hr beyond 5 GiB free (paused-disk$0 compute while paused; disk $0.000108/GiB-hr beyond 5 GiB free (paused-disk billing not stated explicitly) | Pause stops compute and keeps memory and durable filesystem (open TCP/SSH/previewPause stops compute and keeps memory and durable filesystem (open TCP/SSH/preview connections close); configurable max duration and pause retention; sticky sessions run until terminated; snapshots of disk and memory for forks; disk wiped on teardown, except sessions Tenki terminates keep their disk for 7 days | VM | May 2026 | Start under a second from custom images mounted straight from Ceph; root, Docker, systemd; checkpoints and forksStart under a second from custom images mounted straight from Ceph; root, Docker, systemd; checkpoints and forks; persistent volumes; preview URLs per port; secrets swapped in outside the sandbox; TypeScript, Python, Go SDKs and CLI; own hardware, storage and network; also sells GitHub Actions runners and an AI code reviewer; parent company Luxor Technology; startup program up to $50K credits |
| Core | Core | $0.20 | $0 compute; snapshots $0.07/GB-month | Suspended sandboxes pay snapshot storage onlySuspended sandboxes pay snapshot storage only; while running, idle still bills 20% of allocated vCPU | VM | — | Free network egress |
| Core | Core | EUR only, flat monthly, no metered compute: Base €6/mo for 1 persistent terminal (2 vCPUEUR only, flat monthly, no metered compute: Base €6/mo for 1 persistent terminal (2 vCPU, 4 GB RAM, 50 GB total storage quota) = €0.0082/hr (€6/month ÷ 730); Pro €28/mo for 5 terminals (2 vCPU/4 GB each, 100 GB shared); Max €50/mo for 10 terminals (250 GB shared); free SSH client (5 GB); connecting your own machine is free | Same as running (flat €6/mo per terminal; sessions keep running) | Paid cloud terminals keep running when detached (tmux session, files persist)Paid cloud terminals keep running when detached (tmux session, files persist); free SSH-access client pauses after 6 hours detached and wakes on reconnect | Container | — | Browser workspace for coding-agent terminals on your own Mac/Linux machine or Terminalbox CloudBrowser workspace for coding-agent terminals on your own Mac/Linux machine or Terminalbox Cloud; templates for Claude Code, GitHub Copilot CLI, Codex, Gemini CLI, OpenCode, Hermes, OpenClaw; tmux; TBX CLI attach; hosted in Germany (fra1), one production region; no SLA, no restore guarantee, operators can see terminal traffic (security page); 14-day first-payment refund |
| Core | Core | $0.0123 | Same as running (flat monthly plan) | Always on 24/7, billed monthly; no auto-suspendAlways on 24/7, billed monthly; no auto-suspend; you can stop and start it from the app (disk kept, processes end); provider backups can be restored | — | Aug 2026 | Always-on Linux computer for coding agents (Claude Code, Codex, OpenCode, Cursor Agent, Grok BuildAlways-on Linux computer for coding agents (Claude Code, Codex, OpenCode, Cursor Agent, Grok Build, Pi) with your own subscriptions; persistent agent sessions; web terminal and SSH for desktop harnesses; Remote MCP to drive the computer from a laptop agent; Squad multi-agent orchestration and Squad Memory; scheduled agents; Agent Browser (real Chrome) included; iOS Simulator and Cloud Mac through NoMac; workload protection against runaway memory use; locations in America, Europe and Asia/Oceania; bring your own Ubuntu server |
| Core | Core | $0.02 | $0 RAM while asleep; disk $0.25/GB-month | By default a server scales to zero after about 10 minutes without incoming traffic (RAMBy default a server scales to zero after about 10 minutes without incoming traffic (RAM freed, disk kept) and wakes on the next request in 1-3 s; sleep and wake on demand through the API; never_sleep (verified accounts) keeps it always on | gVisor | Oct 2026 | One POST /create with no signup returns root SSH to an Ubuntu box, a live HTTPS subdomain and an api_keyOne POST /create with no signup returns root SSH to an Ubuntu box, a live HTTPS subdomain and an api_key; SKILL.md and llms.txt written for agents; custom domains on verified accounts (up to 3 per server); tenants network-isolated from each other; outbound SMTP blocked; abuse suspended automatically; runs on the operator's own servers in the Czech Republic (EU, GDPR); markets itself as hosting for vibe-coded apps and as a sandbox and persistent workspace for AI agents |
| Core | Core | Containers billed per hour (no shorter interval)Containers billed per hour (no shorter interval), dedicated vCPUs never preempted: $0.06/vCPU-hour + $0.01/GB-hour memory; persistent volumes $0.000137785/GB-hour of provisioned capacity until deleted; GPUs hourly with a monthly commitment: H200 $5.90, B200 $8.12, B300 $9.23. Calculator options: 2-64 vCPU, memory 8-768 GB, so there is no 4 GB option (2 vCPU/8 GB = $0.20/h). Bunker pricing is not published. Chat $20/month | Volumes bill on provisioned GB-hours until deleted | Bunkers keep home and /workspace on an encrypted workspace volume across restartsBunkers keep home and /workspace on an encrypted workspace volume across restarts; Containers are Running or Stopped (Stopped keeps the saved configuration and attached volumes; Deploy boots a new enclave) | VM | — | Private AI company (encrypted chat, private inference API, Containers, Bunkers)Private AI company (encrypted chat, private inference API, Containers, Bunkers); a Bunker is "a persistent, hardened confidential cloud computer for you or your agents": Ubuntu, passwordless sudo, Docker, encrypted /workspace of 2-64 GB, preinstalled coding agents (Claude Code, Codex, OpenCode, Crush, Qwen Code, pi, Goose, Droid, GitHub Copilot), web terminal, SSH with attestation-pinned keys, remote desktop; a Perimeter (second enclave) holds credentials and sets egress policy; quota tiers of 5, 25 or 50 Containers plus as many Bunkers |
| Core | Core | $0.1488 | Hibernated snapshot storage price not published | Hibernate snapshots disk and memory, resume in 500 ms (P95)Hibernate snapshots disk and memory, resume in 500 ms (P95); fork from a hibernated sandbox; legacy SDK hibernates automatically after a timeout once the session disconnects | VM | Dec 2024 | 2.7 s cold start (P95), 500 ms snapshot resume (P95), VM cloning under 1 s2.7 s cold start (P95), 500 ms snapshot resume (P95), VM cloning under 1 s; 2-64 vCPUs and 1-128 GB RAM (product page) vs 0.1-16 cores and 1-8 GB per CPU, default 1 vCPU / 2 GiB (docs); snapshots built from a Docker image or Dockerfile; Python and TypeScript SDKs and CLI (together-sandbox); suggested for agent tool calls and RL rollouts; codesandbox.io pages blocked automated fetches |
| Core | Core | CNY only, per second while running: vCPU ¥0.00003/vCPU-s (¥0.108/vCPU-h)CNY only, per second while running: vCPU ¥0.00003/vCPU-s (¥0.108/vCPU-h), memory ¥0.000015/GiB-s (¥0.054/GiB-h); 10 GB local disk free; persistent storage (templates, snapshots, paused sandboxes) 60 GB free then ¥0.0005/GB-h; new accounts get 100 CPU-hours + 200 GiB-hours free for 30 days. 2 vCPU / 4 GiB = ¥0.432/h. No USD price | Paused sandbox billed as persistent storage: 60 GB free, then ¥0.0005/GB-h | Default timeout 5 minutes, adjustable with set_timeoutDefault timeout 5 minutes, adjustable with set_timeout; at timeout the sandbox is recycled, or paused if auto-pause is enabled; pause keeps filesystem and memory (running processes) and resume restores them; 24-hour continuous running limit, reset by a pause; paused sandboxes count as persistent storage | Container | — | E2B-compatible mode (use the E2B SDK); Python SDK and CLIE2B-compatible mode (use the E2B SDK); Python SDK and CLI; custom templates from snapshots, with examples for Claude Code; public URLs; internet access control; regions North China (Ulanqab, default), East China (Shanghai), Vietnam (Ho Chi Minh City), US (Los Angeles), US (Denver); up to 8 vCPU / 8 GB, CPU:memory ratios 1:1, 1:2, 1:4; 20 concurrent sandboxes; 5 MB/s public bandwidth |
| Core | Core | Plan quotas, no per-resource hourly rate published: Hobby $0/mo (2 running instancesPlan quotas, no per-resource hourly rate published: Hobby $0/mo (2 running instances, 2 vCPU and 4 GiB total, 5 standby, 8 GiB standby storage, 10 GiB image storage); Team $39/mo (8 running, 8 vCPU, 8 GiB, 50 standby, 15 GiB standby storage); Pro $199/mo (24 running, 24 vCPU, 32 GiB, 5,000 standby, 1 TiB standby storage); Enterprise custom (dedicated hosts, on-prem, BYOC); quotas are account-wide, not sandbox sizes | $0 compute (scale to zero), within plan standby quotas | Scale to zero: instance state is snapshotted and resumed in milliseconds on the nextScale to zero: instance state is snapshotted and resumed in milliseconds on the next request; standby instances counted per plan; checkpoints and forking | VM | Apr 2024 | <10 ms cold start claim; scale to zero and resume in ms; checkpoints, forking, templates<10 ms cold start claim; scale to zero and resume in ms; checkpoints, forking, templates; claims 100K+ scale-to-zero agents per server; on-prem, BYOC or dedicated hosts; runs on bare metal, EC2/GCE/Azure VMs or as a node in EKS/GKE/AKS; built on the open-source Unikraft project |
| Core | Core | Monthly plans (EU prices + VAT; Americas higher): Free $0 (sites only, no computer)Monthly plans (EU prices + VAT; Americas higher): Free $0 (sites only, no computer); Small $5/mo always-on 1 GB / 1 vCPU, 15 GB disk (Americas $10); Plus $20/mo 4 GB / 4 vCPU, 40 GB disk (Americas $30); Pro $70/mo 16 GB / 16 vCPU (Americas $90); Max $200/mo 48 GB / 32 vCPU; included boosts (1 h at up to twice the size); Uno AI add-on +$15 to +$100/mo; no per-GB traffic charge. No 2 vCPU / 4 GiB plan; nearest is Plus 4 vCPU / 4 GB at $20/mo (about $0.0274/h) | Included in the monthly plan (sleeping computers don't count toward plan memory) | A computer naps after 10 minutes idle (economy mode) and wakes in about a second on SSHA computer naps after 10 minutes idle (economy mode) and wakes in about a second on SSH, HTTPS, the Telegram relay, a run or Uno Work; restores from a memory snapshot (~3 ms VM restore); sleeping computers don't count toward the plan's memory; the 3-day no-card trial computer sleeps after 5 minutes idle; machines can be created with an expiry (expires_in_s, ttl) | VM | Oct 2026 | Root SSH, Docker, any port; Claude Code, Codex, tmux and node preinstalled, room for 3-4 agents on PlusRoot SSH, Docker, any port; Claude Code, Codex, tmux and node preinstalled, room for 3-4 agents on Plus; resize memory without rebuild, clone, snapshot; REST API, CLI and MCP; agents can register unclaimed accounts for a person to claim; git push and GitHub App deploys with rollback; free static sites; OpenAI-compatible LLM API; S3 storage, Postgres, cron, secrets; GPU by the hour; computers run in Amsterdam (Americas accounts go online from a US exit IP; a New York region is listed as fully booked); minimum computer 512 MB / 1 vCPU / 5 GB disk; Uno4 Labs LLC; card or crypto payment |
| Core | Core | Free $0 (1 concurrent container, 1 request per 42 s, 15-min TTL max, execute only)Free $0 (1 concurrent container, 1 request per 42 s, 15-min TTL max, execute only); Development $7/mo (1 vCPU, 2 GB RAM, 1 concurrent execution, 7 requests/min, sessions and services); Production $91/mo (4 vCPU, 8 GB, 4 concurrent, 91 requests/min); Business $203/mo (8 vCPU, 16 GB, 8 concurrent, 203 requests/min); custom tiers in $7/month steps (tier N = $7N/mo and 7N requests/min); 7-27% discounts for 3-12 months prepaid; pay with XMR, BTC, LTC, DOGE or card; no 2 vCPU/4 GB plan, and tiers are rate limits rather than always-on boxes | Flat monthly tiers; homepage says freezing means you pay only for active compute | /execute and /run TTL 1-900 s (default 60 s); Free tier TTL max 15 min/execute and /run TTL 1-900 s (default 60 s); Free tier TTL max 15 min; sessions and services can be frozen to keep their state and unfrozen later, and services auto-unfreeze on an HTTP request; services 'deep freeze' when the API key expires and keep their data; snapshots (standard ones pause briefly, hot ones don't) | Container | Nov 2025 | REST API for code execution in 42+ languages; execute, interactive sessions (shell or REPL, tmux/screen)REST API for code execution in 42+ languages; execute, interactive sessions (shell or REPL, tmux/screen), persistent services with automatic HTTPS on *.on.unsandbox.com and custom domains, snapshots and clones; zerotrust (no internet, default) or semitrusted (egress proxy) network modes; 1-8 vCPU per execution with 2 GB RAM per vCPU; SDKs in Python, JavaScript, Go, Ruby, Rust and Bash plus 42 language clients; 'un' CLI; anonymous sign-up and crypto payments; machine-readable pricing JSON for agents |
| Core | Core | $0.0111 | $0 CPU; storage $0.10/GB-month | Standard boxes freeze after inactivity, keep storage and bill no CPU while idleStandard boxes freeze after inactivity, keep storage and bill no CPU while idle; keep-alive (Fixed) boxes bill continuously | — | — | Sizes Small 2/4, Medium 4/8, Large 8/16 |
| Core | Core | $0.0884 | $0 CPU and memory; disk R3/GB-month after 5 GB free (about $0.18/GB-month) | Pause keeps state and bills only the diskPause keeps state and bills only the disk; running sandboxes are paused when credits run out; files, packages and browser logins live on persistent disks that survive restarts; always-on plan for a small server that never sleeps; no idle auto-pause timer stated | VM | — | Templates boot into a running Next.js, FastAPI or Django appTemplates boot into a running Next.js, FastAPI or Django app; browser sandbox with a real Chromium you sign in to over noVNC, which agents reuse through MCP (browse_url, screenshot_url with session=true); full Linux desktop with Chrome over noVNC; custom templates from your own image; public HTTPS URL, SSH and file APIs per sandbox; hosted MCP server at api.ventaw.com/mcp; Python SDK; network blocks other sandboxes, private networks and cloud metadata; runs in South Africa; payments through Paystack; a Mmogo Media product |
| Core | Core | $0.3408 | $0 compute; snapshots $0.08/GB-month | Persistent sandboxes snapshot the filesystem on stopPersistent sandboxes snapshot the filesystem on stop; max session 24 h on Pro (45 min Hobby); snapshots expire 30 days after last use by default | VM | — | 19 regions; Drives (persistent storage) in public beta; up to 8 vCPU / 16 GB on Pro, 32 / 64 on Enterprise19 regions; Drives (persistent storage) in public beta; up to 8 vCPU / 16 GB on Pro, 32 / 64 on Enterprise; 64 GB NVMe disk |
| Core | Core | $0.1325 | $0 compute when snapshotted; storage billed (₹0 during beta) | No maximum run time; with autoSleep, a sandbox idle for 1 minute (default) is snapshottedNo maximum run time; with autoSleep, a sandbox idle for 1 minute (default) is snapshotted to disk with memory and wakes in under a second on the next request or public-URL hit (same files and processes); snapshotted sandboxes auto-deleted after 7 days (default); manual sleep and wake API | VM | Feb 2026 | Sandbox ready in under 1 s claim; code interpreter, shell, PTY, background commandsSandbox ready in under 1 s claim; code interpreter, shell, PTY, background commands, file system and file watching APIs; custom Docker images; Docker and K3s inside sandboxes; internet allowlists and blocklists; Tailscale VPN; public URLs; VNC and SSH; MCP integration; TypeScript and Python SDKs, REST API, CLI; self-hosted deployment documented (repo is public but licensed as proprietary); Sappal Cloud Services Private Limited (India) |
| Core | Core | $0.1026 | $0 compute while paused (filesystem snapshot kept$0 compute while paused (filesystem snapshot kept; its storage charge not stated) | Lifetime configurable per instance, at most 24 hours per sandbox instance (AgentKitLifetime configurable per instance, at most 24 hours per sandbox instance (AgentKit limits page, not adjustable); pause releases CPU and memory and stops compute billing, keeps the filesystem as a snapshot, but memory, processes and connections are not kept | Container | Oct 2025 | ByteDance's agent platform: agent runtime, gateway for models and MCP tools, identity and credentials, Viking memoryByteDance's agent platform: agent runtime, gateway for models and MCP tools, identity and credentials, Viking memory, knowledge bases, skills and evals; sandbox templates on veFaaS (Code, Browser, Computer, All-in-One, ArkClaw and Hermes sandboxes); E2B-compatible API; regions North China 2 (Beijing) and East China 2 (Shanghai), plus the BytePlus edition outside China; 50 sandbox templates per account and region; 10 GB max imported image |
| Core | Core | Billed per second at the macOS 26 runner rate: warp-macos-26-arm64-6x (6 vCPU, 14 GBBilled per second at the macOS 26 runner rate: warp-macos-26-arm64-6x (6 vCPU, 14 GB, 60 GB disk, default) $0.08/min ($4.80/h); warp-macos-26-arm64-12x (12 vCPU, 28 GB) $0.16/min ($9.60/h). Paused: no compute charge; saved disk at cache pricing $0.20/GB-month plus $0.0001 per save or restore. No 2 vCPU/4 GiB size | $0 compute while paused; saved disk $0.20/GB-month plus $0.0001 per cache$0 compute while paused; saved disk $0.20/GB-month plus $0.0001 per cache operation | Pause saves the VM's disk (compressed into WarpBuild cache) and stops itPause saves the VM's disk (compressed into WarpBuild cache) and stops it; resume starts a new VM from that disk, keeping the home folder and Homebrew (/opt/homebrew) but not ~/Library, /tmp, /usr/local, /Applications, running processes or memory; a sandbox expires 7 days after creation, running or paused; kill deletes it without saving | VM | Oct 2026 | macOS sandboxes from the wb CLI (macOS or Linux) with a CI-scoped API key: exec (foreground or background)macOS sandboxes from the wb CLI (macOS or Linux) with a CI-scoped API key: exec (foreground or background), process list and stdin, interactive PTY, file ls/cat/put/stat, metrics, env, VNC desktop (wb sbx vnc); same full macOS 26 image as WarpBuild's macOS runners with Xcode preinstalled; WarpBuild's main product is GitHub Actions runners (Linux, macOS, Windows; managed or bring-your-own-cloud) |
| Core | Core | EUR only, per second while running on requested resources: €0.000020/vCPU-sEUR only, per second while running on requested resources: €0.000020/vCPU-s (€0.072/vCPU-h) + €0.000004/GiB-s (€0.0144/GiB-h); retained disk €0.10/GiB-month, checkpoints and templates €0.03/GiB-month; plans Basic free (template shape only, 10 GiB disk, 20 concurrent), Pro €150/month (CPU/RAM overrides, 20 GiB disk, 100 concurrent), Enterprise custom; 2 vCPU / 4 GiB = €0.2016/h (needs Pro for a custom shape). No USD price | €0.10/GiB-month retained disk (paused or stopped)€0.10/GiB-month retained disk (paused or stopped); machine-state image not metered separately | Timeout per sandbox (max 24 hours on Basic and Pro); default on_timeout is killTimeout per sandbox (max 24 hours on Basic and Pro); default on_timeout is kill; with on_timeout pause the machine is frozen with processes and memory and later resumed in place (falls back to a disk restart when no machine-state image exists); optional auto_resume on the next file or process request; paused, stopped, failed and lost sandboxes' disks expire after 7 days; named disk checkpoints branch new sandboxes | — | Jun 2026 | EU-hosted (EU data centers); SSH with one command (ssh base@box.watasu.io) or the CLI; HTTP API with idempotency keysEU-hosted (EU data centers); SSH with one command (ssh base@box.watasu.io) or the CLI; HTTP API with idempotency keys; HTTPS preview URLs per exposed port; egress controls; templates with build logs; named persistent volumes; Python, TypeScript and Rust SDKs; Watasu MCP can run sandbox commands; 1, 2, 4, 6 or 8 vCPU and 512 MiB-8 GiB memory; part of Watasu's EU app-hosting platform (apps, add-ons, Postgres) |
| Core | Core | $0.128 | N/a (deleted after the task; no paused state) | Disposable: when the task ends the machine and everything on it is deletedDisposable: when the task ends the machine and everything on it is deleted; maximum lifetime per sandbox 6 h (Pay as you go), 4 h (Solo), 24 h (Pro), 48 h (Team); no pause or suspend stated | VM | — | 'Give your AI a computer of its own': connect Claude (claude.ai, desktop, mobile) or ChatGPT as a custom MCP connector'Give your AI a computer of its own': connect Claude (claude.ai, desktop, mobile) or ChatGPT as a custom MCP connector, or Claude Code, Cursor, VS Code Copilot, Codex CLI, Gemini CLI; MCP over streamable HTTP and stdio, Python SDK, REST API, LangChain Deep Agents backend; boots from microVM memory snapshots with warm pools; background jobs, batches, Budget Guard, execution receipts and artifacts; custom templates on Pro; runs on refurbished servers in the EU; dedicated servers for Enterprise; KXCOMM s.r.o. (Czech) |
| Core | Core | $0.0333 | $0 runtime while stopped (disk and memory kept; no storage charge stated) | Stop saves disk and memory, including running programsStop saves disk and memory, including running programs, and start resumes from the same moment; a computer stops on its own after 1 hour of running (each start allows up to 1 h, which they plan to raise) or after an hour of near-zero CPU; stopped full VMs are not charged for runtime; rm deletes with no backup | VM | — | Ready in about 190-200 ms (claim; starts from a saved, already-booted machine)Ready in about 190-200 ms (claim; starts from a saved, already-booted machine); fork copies a running computer including memory and programs; Debian 13 with sudo, Python 3.12, Node.js 24, headless browser, Claude Code and Codex preinstalled; computer use; shared volumes; Vault (early access) keeps API keys outside the VM; CLI with JSON output and SDKs; up to 4 running VMs per account; runs on its own servers in Delhi, India (slower from Europe and the US); requests queue when capacity is full |
| Core | Core | Monthly plans shared with Turborg, xShellz's chat-bot product (annual = 10x monthly)Monthly plans shared with Turborg, xShellz's chat-bot product (annual = 10x monthly): Free $0: 1 box, 0.5 core, 512 MB, 0.3 GB disk; Starter $10/mo: 1 box, 1 core, 1 GB, 10 GB; Pro $24/mo: 2 boxes, 2 cores, 2 GB, 25 GB; Max $60/mo: 5 boxes, 4 cores, 4 GB, 50 GB. No 2 vCPU / 4 GiB box: nearest are Pro (2 cores, 2 GB, $24/mo) and Max (4 cores, 4 GB, $60/mo); the table doesn't say whether Pro and Max resources are per box or per plan | Flat monthly plan price (boxes are always on) | Boxes run 24/7 with no idle shutdown (paid plans 'always-on')Boxes run 24/7 with no idle shutdown (paid plans 'always-on'); /root is a persistent volume kept across restarts, resumes and image updates, the rest of the filesystem resets when the box is rebuilt on a newer image; abandoned boxes are garbage-collected; the SDK also makes throwaway boxes (Sandbox.create ... kill) or keeps one by name (get_or_create). The Free plan's uptime is stated two ways: '30-day trial, then on-demand' in the plan table, but 24/7 with no trial countdown and no idle shutdown in the FAQ | gVisor | Jul 2026 | Hardened Ubuntu box you hand an agent root on, over SSH or a browser terminal; apt install anythingHardened Ubuntu box you hand an agent root on, over SSH or a browser terminal; apt install anything; six coding agents preinstalled (borg, Claude Code, Codex, Gemini CLI, opencode, pi) plus OpenClaw and Hermes, crewai and langgraph; MCP server so Claude Code drives the box over SSH; QA kit (promptfoo, pytest, ruff, shellcheck, uv); live CPU, memory, disk, network and process monitoring; one-click install of 23 apps (WordPress, Ghost, Hugo, Next.js...) with publishing to a subdomain; xshellz.box manifest reinstalls packages on boot; MIT SDKs for Python, TypeScript, Go, Java, Ruby, PHP and Rust (1 star each); separate 'AI Sandbox Shell' product marked coming soon; xShellz has run IRC shells since 2009 |
| Core | Core | Billed in Yeps (1 Yep = 1 second of execution, rounded up): Developer free (50K YepsBilled in Yeps (1 Yep = 1 second of execution, rounded up): Developer free (50K Yeps, 300 executions a day, 1 concurrent, limited CPU and memory); Starter $119/mo (15M Yeps, 300 MB RAM sandboxes, extra Yeps $0.00001 each); Growth $599/mo (150M Yeps, 600 MB RAM, extra Yeps $0.000005 each); Enterprise from $1,800/mo with customizable CPU and memory. No 2 vCPU / 4 GiB size | $0 (billed per second of execution) | One sandbox per execution; maximum execution time 30 s (Developer), 1 hour (Starter)One sandbox per execution; maximum execution time 30 s (Developer), 1 hour (Starter), 12 hours (Growth); local disk 5 MB, 250 MB or 5 GB by plan | — | Feb 2025 | SDKs: npm @yepcode/run and PyPI yepcode-run; runs Node.js or Python and installs any npm or PyPI package on demandSDKs: npm @yepcode/run and PyPI yepcode-run; runs Node.js or Python and installs any npm or PyPI package on demand; secrets, logs and audit trail; MCP server; on-demand, webhook and scheduled executions; on-premise executors; homepage banner: 'We're not taking on new customers at this time'; docs include a 'Migrating from YepCode Cloud' guide for running the workspace yourself with the CLI; the YepCode Run page (yepcode.io/run) returns 404 |
| Core | Core | $0.0137 | $0 while paused (disk kept) | Runs until you stop it; nothing times outRuns until you stop it; nothing times out; pause saves the disk and stops the meter, and restore boots the same disk (memory is not kept); delete keeps the machine's name, email and wallets on hold | VM | Jul 2026 | Cloud computer for AI coding agents with a browser desktop (ZipOS) and terminalCloud computer for AI coding agents with a browser desktop (ZipOS) and terminal; boots with the Pi agent and has setup guides for Claude, Codex, Cursor, OpenCode, Amp, Grok and others; root access; SSH to name.zbox.sh over IPv6; pre-warmed machines, 'live in seconds'; fork a box; Tailscale private mode; each agent gets an email inbox, public domain, and EVM and Solana wallets; API with TypeScript, Python, Go and Rust SDKs, CLI and hosted MCP (since 2026-10-21); mobile app; run by Tribes Inc. |
| Core | Core | $0.0175 | $0 compute; storage $0.10/GB-month after the first 48 h | Stop hibernates: memory and disk saved, processes resume on startStop hibernates: memory and disk saved, processes resume on start; homepage says idle hibernate stops billing in 60 s (docs give no idle timeout); no cap on run length; credit exhaustion stops (hibernates) sandboxes | VM | Nov 2026 | Live fork of a running VM in about 20 s (processes, memory, disk); snapshots of disk or disk+memoryLive fork of a running VM in about 20 s (processes, memory, disk); snapshots of disk or disk+memory; Linux desktop (paid plans) and Chrome over CDP; SSH and exec; HTTPS port previews; secret vault keeps real keys on host; flight recorder, cost per PR; CI runners (runs-on: zode); TypeScript and Python SDKs, CLI, MCP; region EU (GCP europe-west1), US-East soon; ready in about 12 s |
| Runtime / SDK | Runtime / SDK | Free, Apache-2.0; runs on your own Kubernetes clusterPricing page | — | operatingMode Suspended terminates the pod but keeps the Sandbox object and its volumesoperatingMode Suspended terminates the pod but keeps the Sandbox object and its volumes, and resume creates a new pod that reattaches them; shutdownTime expiry deletes pod and service (shutdownPolicy Retain or Delete); SandboxWarmPool keeps pre-warmed sandboxes | gVisor | Aug 2025 | Sandbox CRD and controller under Kubernetes SIG Apps; stable identity and hostname, persistent storageSandbox CRD and controller under Kubernetes SIG Apps; stable identity and hostname, persistent storage; SandboxTemplate, SandboxClaim and SandboxWarmPool extensions; sandbox router for SDK traffic; Go and Python SDKs; Kubernetes 1.25+ (1.29+ recommended); 4,610 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free open-source software (Apache-2.0); you pay for your own Kubernetes infrastructurePricing page | Self-hosted; idle actors are snapshotted out of memory so they use no workerSelf-hosted; idle actors are snapshotted out of memory so they use no worker compute | Snapshots idle actors out of memory (RAM and filesystem captured) and restores them onSnapshots idle actors out of memory (RAM and filesystem captured) and restores them on request onto a small pool of warm worker pods; resume under 500 ms, 500+ activations per second (README claims); 10x-30x oversubscription | gVisor | May 2026 | Runs on standard Kubernetes (Pods, autoscaling, NetworkPolicies); per-actor SPIFFE X.509 identityRuns on standard Kubernetes (Pods, autoscaling, NetworkPolicies); per-actor SPIFFE X.509 identity; zero credentials in the sandbox, egress gateway injects credentials outside; default-deny networking; pluggable egress policy point; hosts OCI images for ADK, LangChain, Claude Code, Codex, Antigravity, MCP servers; used by Google's Agent Executor and kagent; originally created by Google; 6,641 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | N/a (self-hosted) | Pause/resume, snapshot, and scale to zero on idle or timeout (idle timeout set perPause/resume, snapshot, and scale to zero on idle or timeout (idle timeout set per sandbox) | Container | Dec 2025 | Self-hosted on Kubernetes 1.28+ with one kubectl apply; REST API, MCP server and built-in web UISelf-hosted on Kubernetes 1.28+ with one kubectl apply; REST API, MCP server and built-in web UI; full E2B protocol and SDK compatibility; code execution, browser, desktop (VNC) and shell sandboxes based on AIO Sandbox; pre-warmed sandbox pools; templates and blueprints; multi-tenant; leader election for HA; 232 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs on your own machine or your own GCPFree, open source (MIT); runs on your own machine or your own GCP, AWS or Kubernetes account | N/a (self-hosted) | Desktops persist until deleted (Desktop.delete or agentdesk delete) | Container | Jan 2024 | Python SDK and CLI to create, view and control desktops (open URL, mouse, keyboard, screenshots)Python SDK and CLI to create, view and control desktops (open URL, mouse, keyboard, screenshots); built on agentd, a daemon with a REST API inside the desktop; implements the DeviceBay protocol; grid processor for screenshots; 231 GitHub stars; agentsea.ai returns 404 |
| Runtime / SDK | Runtime / SDK | Free, MIT, self-hostedPricing page | — | Snapshot-backed: pause <100 ms releases CPU and memory, boot or resume <50 msSnapshot-backed: pause <100 ms releases CPU and memory, boot or resume <50 ms; incremental memory and filesystem snapshots persisted to S3-compatible storage or a shared filesystem | VM | Jul 2026 | Powers agentic RL training for Kimi K3; overlaybd on-demand OCI image loading (1.5 million images in production)Powers agentic RL training for Kimi K3; overlaybd on-demand OCI image loading (1.5 million images in production); fork a running environment; memory ballooning (9.6x overcommit in production); E2B-compatible HTTP API, use E2B Python/TS SDKs; single node, Docker Compose or Kubernetes; Linux 6.8+ with KVM, or PVM; 3,747 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0)Pricing page | N/a (self-hosted) | Persistent sessions: the agent keeps running after the terminal closesPersistent sessions: the agent keeps running after the terminal closes; per-task timeouts (e.g. 2h); restart-safe session reconciliation | VM | Apr 2026 | Self-hosted runtime for persistent autonomous coding agents; management server with gRPC, WebSocket and HTTPSelf-hosted runtime for persistent autonomous coding agents; management server with gRPC, WebSocket and HTTP; web dashboard (AIWG Cockpit), CLI, REST API; A2A-protocol executor; virtiofs shared storage; per-VM CPU/memory/disk quotas; human-in-the-loop prompt detection; part of the AIWG suite; 14 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); hosted remote backends bill at their own ratesPricing page | N/a (self-hosted; hosted remote backends bill at their own rates) | Agentkernel run creates a temporary sandbox and removes it afterwards (--keep keeps it)Agentkernel run creates a temporary sandbox and removes it afterwards (--keep keeps it); named persistent sandboxes with start/stop/remove and optional --ttl auto-expiry; filesystem snapshots; microVM-only full-state snapshots keep memory and running processes and can be forked; sessions tie a sandbox to an agent conversation; gc removes expired or stopped sandboxes | VM | Jan 2026 | Rust CLI, HTTP API (OpenAPI), MCP server and desktop app; SDKs for Node.js, Python, Go, Rust and SwiftRust CLI, HTTP API (OpenAPI), MCP server and desktop app; SDKs for Node.js, Python, Go, Rust and Swift; claims sub-125 ms boot and 195 ms from a warm pool on Linux; plugins for Claude Code, Codex, Gemini CLI, OpenCode and any MCP agent; host-side proxy injects secrets so the guest sees only placeholders; permissive, moderate and restrictive security profiles; certificate-authenticated SSH; 18+ templates; pipelines and parallel jobs; signed execution receipts; per-tenant LLM model allowlists; Helm chart and Nomad pack with warm pools and CRDs; by Paul Thrasher; 62 GitHub stars |
| Runtime / SDK | Runtime / SDK | agentOS library free (Apache-2.0). Rivet Cloud: control plane Free $0, Hobby $18/moagentOS library free (Apache-2.0). Rivet Cloud: control plane Free $0, Hobby $18/mo, Team $156/mo, Enterprise $740/mo, plus Actor Nodes per month: R-5 1/4 vCPU/1 GB $18, R-10 1/2 vCPU/2 GB $29, R-20 1 vCPU/4 GB $52, R-40 4 vCPU/8 GB $148, R-160 8 vCPU/16 GB $327, on Google Cloud; no 2 vCPU/4 GB node; YC and a16z Speedrun companies 50% off Team for 12 months | — | As a library the embedding app decides; a VM hosted on Rivet sleeps when idle unless aAs a library the embedding app decides; a VM hosted on Rivet sleeps when idle unless a process is running and starts lazily on the next action; durable files (SQLite-backed root), configuration and cron jobs survive sleep, processes do not | WebAssembly | Apr 2026 | npm library that runs inside your backend; Node.js on native V8, Python and Bash via WebAssemblynpm library that runs inside your backend; Node.js on native V8, Python and Bash via WebAssembly; POSIX filesystem with S3, Google Drive and host-directory mounts; host functions without exposing secrets; software registry (git, curl, ripgrep, DuckDB, Browserbase browser); 4.8 ms p50 cold start and ~22 MB per instance (Rivet's own benchmark); local, BYOC or self-hosted; rivet-dev/agentos 4,899 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (README says Apache 2.0; GitHub detects no license file)Free, open source (README says Apache 2.0; GitHub detects no license file); self-hosted on a Linux host with KVM | N/a (self-hosted) | Per-request VM with a policy timeout (example policy: 5 minutes)Per-request VM with a policy timeout (example policy: 5 minutes); VM pooling and scale-to-zero listed as coming soon | VM | Sep 2025 | Agent-run CLI with YAML capability policies (CPU, memory, timeout, egress domain allowlist, writable paths)Agent-run CLI with YAML capability policies (CPU, memory, timeout, egress domain allowlist, writable paths), default deny; Go SDK; Prometheus metrics; claims sub-200 ms boot; MVP with 9 commits, all on 2025-09-21; 4 GitHub stars. This input row also named ERA (github.com/BinSquare/ERA), a separate local microVM sandbox (krunvm) that is now archived; its README says it is deprecated in favor of smolvm (the smolvm (smol-machines) row) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2025 | A production-ready runtime framework for agent apps with secure tool sandboxing, Agent-as-a-Service APIsA production-ready runtime framework for agent apps with secure tool sandboxing, Agent-as-a-Service APIs, scalable deployment, full-stack observability, and broad framework compatibility. (881 GitHub stars; Apache-2.0; owner agentscope-ai) |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2026 | Agyn is an open-source Kubernetes-native runtime that moves AI agents like Claude Code and Codex from laptops toAgyn is an open-source Kubernetes-native runtime that moves AI agents like Claude Code and Codex from laptops to company infrastructure with the controls enterprises need. (243 GitHub stars; AGPL-3.0; owner agynio) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2025 | The open-source agent SDK that gives every agent its own VMThe open-source agent SDK that gives every agent its own VM. Each VM is a microVM sandbox where the agent installs software, runs code and uses the GPU. Python, Node.js and Rust. (158 GitHub stars; Apache-2.0; owner brekkylab) |
| Runtime / SDK | Runtime / SDK | Free, Apache-2.0, self-hosted Docker imagePricing page | — | Not stated | Container | Aug 2025 | One container with VNC browser, shell, shared filesystem, VS Code Server, Jupyter and MCP serverOne container with VNC browser, shell, shared filesystem, VS Code Server, Jupyter and MCP server; Python SDK (agent_sandbox) and REST API; port forwarding for previews; optional API-key auth; mainland-China image mirror on Volcano Engine's registry (volces.com); 6,342 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0), self-hosted; no hosted service or pricesPricing page | N/a (self-hosted) | Sandboxes are ephemeral containers; checkpoint() snapshots the container and every execSandboxes are ephemeral containers; checkpoint() snapshots the container and every exec is written to a durable ledger, so resume() reconnects or restores after a full restart; agents can be paused, steered and forked (filesystem and installed packages) | Container | Jun 2026 | Control plane for swarms of sandboxed agents: TypeScript SDK (npm alineo), alineod daemon (HTTP and SSE swarm control)Control plane for swarms of sandboxed agents: TypeScript SDK (npm alineo), alineod daemon (HTTP and SSE swarm control), CLI and alineo-mcp for Claude Code, Claude Desktop and Cursor; agents run Pi inside OpenSandbox; fork a live sandbox into sub-agents with spawnDepth and maxAgents caps; steer() mid-response; SQLite or Postgres ledger adapters; OpenTelemetry hooks; secrets injected at exec time; model agnostic; needs Docker and Bun; GitHub DrejT/alineo 44 stars, last push 2026-11-08 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Node.js in your browser. Just like that. (1184 GitHub stars; MIT; owner macaly) |
| Runtime / SDK | Runtime / SDK | Free; Python package MIT, bundled Rust WASM runtime AGPL-3.0-or-later OR BUSL-1.1Free; Python package MIT, bundled Rust WASM runtime AGPL-3.0-or-later OR BUSL-1.1; runs in your own process | N/a (self-hosted library) | — | WebAssembly | Jan 2026 | 'Capability-based sandboxed runtime for AI agents' by Amla Labs; pip install amla-sandbox (0.2.8)'Capability-based sandboxed runtime for AI agents' by Amla Labs; pip install amla-sandbox (0.2.8); runs JavaScript and shell; agents call only the Python tools you pass in, with constraints (parameter bounds, allowed values, max_calls) and method patterns; framework integrations in examples; repo is a release mirror of a private monorepo; GitHub 347 stars, last push 2026-05-15 |
| Runtime / SDK | Runtime / SDK | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (MIT) | — | Full VM checkpoints uploaded to S3 and forked; no idle auto-suspend stated | VM | Mar 2026 | Rust execution backend for agentic RL rollouts (VeRL, OpenRLHF, TRL)Rust execution backend for agentic RL rollouts (VeRL, OpenRLHF, TRL); snapshot mid-task and fork N attempts with fresh entropy, tokens and credentials ('branch-safe'); per-step trajectory tracer and replay; secret broker and default-deny egress proxy; private Docker daemon in the guest; multi-runner pool with Helm and Prometheus; ARM64/Graviton2; GPU via host-mediated llama.cpp/vLLM sidecar; claims sub-150 ms boot; arbor run-benchmark for SWE-bench; MIT; 14 stars |
| Runtime / SDK | Runtime / SDK | Free, AGPL-3.0, self-hostedPricing page | — | Snapshot-and-restore of a running sandbox (processes and files) for backtrackingSnapshot-and-restore of a running sandbox (processes and files) for backtracking; other lifecycle not stated | VM | Aug 2024 | Ubuntu guest with code execution service and VNC server (Chrome for computer use)Ubuntu guest with code execution service and VNC server (Chrome for computer use); REST API, py-arrakis Python SDK, MCP server; automatic port forwarding; Linux with /dev/kvm only; GCP setup guide; 896 GitHub stars; last push 2025-06-02 |
| Runtime / SDK | Runtime / SDK | Self-hosted runtime free (MIT); Ash Cloud (dashboards, monitoringSelf-hosted runtime free (MIT); Ash Cloud (dashboards, monitoring, optional managed compute) free to start with 2,500 events a month, no compute prices published; Enterprise custom (dedicated compute pools, 99.9% SLA) | N/a (self-hosted) | Sessions pause and resume minutes or days laterSessions pause and resume minutes or days later, even on another machine (workspace state saved to S3 or GCS); sandbox pool with capacity limits, LRU eviction and an idle sweep | Process | Feb 2026 | Deploys Claude Agent SDK agents (a folder with CLAUDE.md, skills and .mcp.json) as REST and SSE APIsDeploys Claude Agent SDK agents (a folder with CLAUDE.md, skills and .mcp.json) as REST and SSE APIs; sessions stored in SQLite or Postgres; coordinator mode routes sessions to registered runners; deploy scripts for EC2 and GCE, or any Docker host; CLI @ash-ai/cli and SDK @ash-ai/sdk; GitHub ash-ai-org/ash-ai, MIT, 35 stars, last push 2026-04-01; Conviction Labs |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0 per the archived repo); self-hosted with Docker ComposeFree, open source (Apache-2.0 per the archived repo); self-hosted with Docker Compose; the README also mentions a 'managed DeepAgent backend' that 'we host', with no price | N/a (self-hosted) | Archived README calls the environments 'isolated, ephemeral'Archived README calls the environments 'isolated, ephemeral'; no timeout or pause documented | Container | Nov 2025 | Sandboxed execution engine for AI agents: Codex workspaces that turn Sentry/Glitchtip errors into merge requestsSandboxed execution engine for AI agents: Codex workspaces that turn Sentry/Glitchtip errors into merge requests, and a LangChain DeepAgents sandbox backend; Python toolkit astraforge-toolkit on PyPI (shell, Python REPL, Playwright and image tools); web app streams logs and diffs; Ollama or OpenAI models; the GitHub repo returned 404 on 2026-11-14 (5 stars in the 2026-02-05 archived copy) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Agent Reference Stack for Kubernetes (kars) - an open source stack from Microsoft for running AI agents safely onAgent Reference Stack for Kubernetes (kars) - an open source stack from Microsoft for running AI agents safely on Kubernetes. Multi-runtime, Foundry-aware, hardened per-agent sandboxes, governed egress, end-to-end encrypted inter-agent mesh. (46 GitHub stars; MIT; owner Azure) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs inside your own JVMPricing page | N/a (self-hosted library) | A Bash instance lives until close(); variablesA Bash instance lives until close(); variables, files and cwd persist across exec calls within one instance and are freed on close; per-exec wall-clock timeoutMs, maxCommands limit and cancel() from another thread | Process | Sep 2026 | Java 17+ SDK on Maven Central (io.github.terseprompts:bashkit4j 0.3.0)Java 17+ SDK on Maven Central (io.github.terseprompts:bashkit4j 0.3.0); native libraries bundled for Windows, Linux and macOS (x86-64 and ARM64); opt-in read-only host mounts; pre-seeded files and env; typed ExecResult; one instance per tenant; Java binding of everruns/bashkit (MIT), native libs built from the author's fork; BashTool LLM/agent layer on the roadmap; 4 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Self-Hostable MicroVMs for your Agents (271 GitHub stars; Apache-2.0; owner sahil-shubham) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2025 | Open-source VMs-as-a-service (809 GitHub stars; MIT; owner stanford-mast) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Sandbox for Agents (100 GitHub stars; MIT; owner vrn21) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0) CLI; you pay Hetzner Cloud directly for each server atFree, open source (Apache-2.0) CLI; you pay Hetzner Cloud directly for each server at Hetzner's hourly rates (the demo uses type cx23 in fsn1); buildbox price lists Hetzner's prices by server type and location | N/a (no suspend; Hetzner bills the server until it is deleted) | No suspend: a box runs until you delete it with buildbox deleteNo suspend: a box runs until you delete it with buildbox delete, and Hetzner bills it by the hour until then | VM | Nov 2026 | Go CLI for macOS and Linux (go install); needs Go 1.27+, Tailscale and a Hetzner Cloud API tokenGo CLI for macOS and Linux (go install); needs Go 1.27+, Tailscale and a Hetzner Cloud API token; a box joins your tailnet tagged tag:buildbox with a single-use auth key and is reached only over Tailscale SSH (OpenSSH off); installs OpenCode, Claude Code, Codex or Grok Build, not signed in, plus Git, a C toolchain and Python 3; buildbox forward brings box ports to localhost; ready in a few minutes (demo: 2 min 36 s); not affiliated with Hetzner or Tailscale; GitHub hkluke/buildbox 4 stars, Apache-2.0; v0.1.1 released 2026-11-13 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Embedded application kernel, inspired by gVisor (221 GitHub stars; Apache-2.0; owner butter-dot-dev) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT) daemon you run on your own Linux hostFree, open source (MIT) daemon you run on your own Linux host; no hosted plan or price is published (sandbox.byggflow.com, the repo's homepage, returns Vercel DEPLOYMENT_NOT_FOUND) | — | No pause: a sandbox lives until sbx rm, or for --ttl seconds after the client disconnectsNo pause: a sandbox lives until sbx rm, or for --ttl seconds after the client disconnects (capped by the daemon's max_ttl); a sandbox can be saved as a template (snapshot) and new sandboxes created from it; templates expire after template_expiry_days | Container | Apr 2026 | Sandboxd daemon with Unix-socket and TCP API, sbx CLI and an in-sandbox guest agentSandboxd daemon with Unix-socket and TCP API, sbx CLI and an in-sandbox guest agent; warm pool of pre-started sandboxes, allocation under 5 ms (README claim); TypeScript, Go and Python SDKs; MCP server (@byggflow/sandbox-mcp); port tunnels; egress rules with HTTPS interception through a per-sandbox CA; multi-tenant identity scoping; Linux and macOS binaries and a Docker image; GitHub byggflow/sandbox 0 stars, MIT; latest release v0.0.5 (2026-04-14), last push 2026-06-04 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); you pay for your own servers and AI provider keys (priceFree, open source (Apache-2.0); you pay for your own servers and AI provider keys (price per hour for 2 vCPU / 4 GiB: n/a (self-hosted)) | N/a (self-hosted) | Self-hosted: the desktop container runs until you stop itSelf-hosted: the desktop container runs until you stop it; a task runs until done or until it asks for help, and a person can take over and resume the agent | Container | Feb 2025 | Self-hosted AI desktop agent on Ubuntu 22.04 with XFCE, Firefox, VS Code and password-manager supportSelf-hosted AI desktop agent on Ubuntu 22.04 with XFCE, Firefox, VS Code and password-manager support; REST API for tasks and desktop control; Docker Compose, Railway or Helm on Kubernetes; Anthropic, OpenAI or Gemini keys; GitHub bytebot-ai/bytebot, 11,071 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); runs in your own process or CLIPricing page | N/a (self-hosted library) | Each task runs once in its own sandbox with CPUEach task runs once in its own sandbox with CPU, memory and timeout limits and automatic retries; no persistent sandbox | WebAssembly | Dec 2025 | 'Secure runtime to sandbox AI agent tasks': @task decorator in Python (pip install capsule-run) or task() in TypeScript'Secure runtime to sandbox AI agent tasks': @task decorator in Python (pip install capsule-run) or task() in TypeScript (@capsule-run/sdk); outbound HTTP blocked unless allowed_hosts lists the domain; file access only to declared or mounted directories; Python and TypeScript adapters for running untrusted code; same org also makes capsulerun/bash (sandboxed bash for agents) and vpod (Linux sandboxes in the browser); v0.8.10 (2026-04-27); GitHub 304 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); runs on your own Linux x86_64 host with /dev/kvmPricing page | N/a (self-hosted) | Ephemeral: a cell is sealed for a run and dissolves when it finishesEphemeral: a cell is sealed for a run and dissolves when it finishes; tool lends are revoked at the end; no pause or persistence | VM | Aug 2026 | Cells borrow digest-pinned tools from container images instead of building a Linux environmentCells borrow digest-pinned tools from container images instead of building a Linux environment; per-call authority (an attested interpreter running model-written code is demoted to the agent lane); no network and only /tmp or /celln/work writable by default; workspace and HTTPS tools go through host brokers with their own grants; `celln agent` one-liner drives a model provider (Anthropic, OpenAI and others); starter packages give a Sympozium multi-agent fleet grep, jq and similar tools; hardware verification tests (`celln verify`); Rust; Homebrew install; no ARM64 build; 62 GitHub stars |
| Runtime / SDK | Runtime / SDK | ClawRun is free and open source (Apache-2.0)ClawRun is free and open source (Apache-2.0); you pay Vercel and your LLM provider. ClawRun's cost page (Vercel prices as of March 2026) says sandboxes default to 2 vCPU / 4 GB and cites Vercel Hobby free tier (5 active-CPU hours and 420 GB-hours of memory per month) or Pro $20/month plus $0.128/active-CPU-hour and $0.0212/GB-hour memory, snapshots $0.08/GB-month | Billed by Vercel: no compute while asleepBilled by Vercel: no compute while asleep; snapshot storage $0.08/GB-month on Vercel Pro, per ClawRun's cost page | Sandbox stays up while active (heartbeats every 60 s extend its TTL)Sandbox stays up while active (heartbeats every 60 s extend its TTL), then snapshots its filesystem and stops when idle (default 10 minutes); wakes from the latest snapshot on a webhook message or a cron schedule; files, config, memory and installed packages persist, running processes don't; keeps the 3 newest snapshots; Vercel caps one wake at 45 minutes on Hobby and 5 hours on Pro | VM | Feb 2026 | CLI (npx clawrun deploy) and web dashboard; deploys ZeroClaw agents (OpenClaw and NanoClaw coming soon) with a Next.jsCLI (npx clawrun deploy) and web dashboard; deploys ZeroClaw agents (OpenClaw and NanoClaw coming soon) with a Next.js orchestrator on your own Vercel account and Upstash Redis for lifecycle state; 15 messaging channels, of which Telegram, Discord, Slack, WhatsApp, Lark and QQ can wake a sleeping sandbox; 48 LLM providers through ZeroClaw; cron-scheduled agents; network policies; TypeScript SDK; GitHub clawrun-sh/clawrun 94 stars, last push 2026-06-12 |
| Runtime / SDK | Runtime / SDK | Free, MIT open source; you supply the hostsPricing page | — | Baked 'warm spore' microVM snapshots (memory and disk) that resume and forkBaked 'warm spore' microVM snapshots (memory and disk) that resume and fork copy-on-write; lifecycle run by SporeVM | VM | Feb 2026 | Compiles cleanroom.yaml policy into SporeVM config, fails closed; deny-by-default network allowlistsCompiles cleanroom.yaml policy into SporeVM config, fails closed; deny-by-default network allowlists; provenance annotations and verify; host-side gateway brokers credentials so secrets stay out of the guest; content cache; fork 100 copies of a warm spore; 70 GitHub stars; first release v0.1.0 2026-02-28 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); you deploy it to your own Cloudflare account and pay CloudflareFree, open source (MIT); you deploy it to your own Cloudflare account and pay Cloudflare for Workers, Containers, R2 and D1 | Cloudflare R2 storage for stopped-run snapshots, billed by Cloudflare | Bounded runs end after verification and return receiptsBounded runs end after verification and return receipts; live runs stay up for ttlSeconds (60 s to 30 days); stop snapshots the workspace to R2 and stops the runner, resume restores the snapshot into a new runner; fork and delete also supported | Container | May 2026 | Fresh Cloudflare Container per task with the repo already cloned; HTTP API (POST /api/runs) and TypeScript adaptersFresh Cloudflare Container per task with the repo already cloned; HTTP API (POST /api/runs) and TypeScript adapters; live steering: exec, read/write files, authenticated dev preview; verification commands, diff, receipts and one named artifact per run; opt-in desktop runner with browser terminal and noVNC Chromium (still 'Proving'); runner instance type set at deploy time; typed workspaces for agent evaluation with rubrics; MIT, 56 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT) library; runs on your own Cloudflare account and is billed atFree, open source (MIT) library; runs on your own Cloudflare account and is billed at Workers, Durable Objects and Containers rates (see the Cloudflare Sandboxes row) | — | The workspace filesystem lives in a Durable Object's SQLite and persistsThe workspace filesystem lives in a Durable Object's SQLite and persists; execution backends connect lazily on first use; container lifecycle follows Cloudflare Containers (see the Cloudflare Sandboxes row) | Container | Jun 2026 | @cloudflare/computer npm package: a virtual filesystem inside a Durable Object with one execution entry point@cloudflare/computer npm package: a virtual filesystem inside a Durable Object with one execution entry point (workspace.runtime.exec) over pluggable backends; computerd FUSE-mounts the workspace inside the container and syncs over capnweb RPC; Codemode host modules for git, Artifacts and containers; egress policies none, all or custom; examples for MCP, pi, TanStack AI and Cloudflare Think agents; core Workspace API stable, container and Worker shell backends experimental; GitHub cloudflare/computer 9,919 stars, MIT; latest release @cloudflare/computer 0.11.2 (2026-11-12) |
| Runtime / SDK | Runtime / SDK | Free, open source (server stack AGPL-3.0; Go, PythonFree, open source (server stack AGPL-3.0; Go, Python, OpenAI and LangChain SDKs Apache-2.0); self-hosted on Linux nodes with KVM | N/a (self-hosted) | Each claim has a TTL (default 5 min, capped at 24 h) after which the node reaps itEach claim has a TTL (default 5 min, capped at 24 h) after which the node reaps it, even if the client vanishes; renew extends it; hibernate snapshots memory and stops the VM, and the next agent access restores it with processes and sessions intact; on_expire=archive hibernates and archives to a checkpoint store instead of destroying, kept for archive_delete_after_seconds or forever; fork, checkpoints and promote | VM | Jul 2026 | Warm-pool claims sub-millisecond, pool miss clones a golden snapshot in tens of msWarm-pool claims sub-millisecond, pool miss clones a golden snapshot in tens of ms, cold boot ~0.2-0.4 s on bare metal (README); per-node sandboxd control plane with memberlist mesh, redirects, draining and live move between nodes; in-guest Rust daemon silkd (exec, files, persistent shells, PTYs, git); Go and Python SDKs, MCP server, OpenAI Agents SDK provider, LangChain toolkit; images base, rt, python, node, browser (headless Chromium over CDP), android (redroid), e2b-rt and e2b-ci (e2b envd for the e2b-compatible data plane); host-side credential injection for egress; Kubernetes via vk-sandbox and sandbox-operator (the sandbox-operator row); underlying cocoon engine (MIT, 266 stars) also runs Windows 11 guests; 89 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2025 | Sandboxes for running code and tools (131 GitHub stars; Apache-2.0; owner nalgeon) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-hostedPricing page | N/a (self-hosted) | Sessions keep state between executions until deleted with DELETE /sessions/{id}Sessions keep state between executions until deleted with DELETE /sessions/{id}; per-request timeout; no idle timeout documented | Container | Dec 2024 | Self-hosted alternative to OpenAI's Code Interpreter; FastAPI service plus an MCP server (Claude Desktop)Self-hosted alternative to OpenAI's Code Interpreter; FastAPI service plus an MCP server (Claude Desktop); pip and conda installs with security controls; plots; host directories can be mounted read-only or read-write; network available by default; README calls it 'A prototype implementation, not intended for production use'; 9 GitHub stars, last push 2025-05-20 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2024 | A simple Python sandbox for helpful LLM data agents (313 GitHub stars; MIT; archived; owner cohere-ai) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Dec 2025 | Run untrusted AI code safely, fast (199 GitHub stars; MIT; owner PwnFunction) |
| Runtime / SDK | Runtime / SDK | Cosmonic Desktop free (public beta, local)Cosmonic Desktop free (public beta, local); Cosmonic Control (Kubernetes) free tier with community support, Enterprise (air-gapped/on-prem support, SLAs) by contact; no published rates; runs on your own machines | $0 compute claimed (scale to zero); you run the hardware | Idle sandboxes scale to zero and start on demand (~0.2 ms per-call instantiationIdle sandboxes scale to zero and start on demand (~0.2 ms per-call instantiation claimed); ~231 KB per idle sandbox claimed | WebAssembly | Sep 2026 | Sandboxes each agent, MCP server, tool and agent-written script in its own Wasm componentSandboxes each agent, MCP server, tool and agent-written script in its own Wasm component; Cosmonic Desktop for macOS, Windows and Linux with an MCP server for Claude Code, Codex, Gemini CLI, opencode and others, plus a built-in app catalog; Cosmonic Control runs the same manifests on EKS, AKS, GKE, OpenShift, on-prem or air-gapped, 'thousands of sandboxes per node'; LLM gateway as a granted capability; open source Agent SDK; Control v0.12 (Kafka capability) |
| Runtime / SDK | Runtime / SDK | Free open source (Apache-2.0); runs on your own Linux host with KVM | — | Sandboxes are ephemeral (a daemon restart tears them down)Sandboxes are ephemeral (a daemon restart tears them down); apps are durable, scale to zero after --idle-timeout and wake on the next request or TCP connection (~125-170 ms snapshot wake); snapshot and fork; persistent volumes | VM | Apr 2026 | Single Go binary (daemon, CLI, TUI, MCP server with 37 tools); forks in ~100-125 ms with lazy userfaultfd memorySingle Go binary (daemon, CLI, TUI, MCP server with 37 tools); forks in ~100-125 ms with lazy userfaultfd memory; boots unmodified OCI images; default-deny egress with hostname allowlist; ingress proxy with automatic HTTPS; Go SDK; pre-1.0 (v0.9.8), README says not yet hardened for untrusted multi-tenant use; 1 GitHub star |
| Runtime / SDK | Runtime / SDK | Free, open source research platform (no license file in the repo)Pricing page | N/a (self-hosted) | — | Container | Aug 2026 | Platform to evaluate, SFT- and RL-train computer-use agents across desktop, browser and mobile; 30k+ verifiable tasksPlatform to evaluate, SFT- and RL-train computer-use agents across desktop, browser and mobile; 30k+ verifiable tasks; 10+ built-in agents (GPT, Claude, Gemini, Qwen, UI-TARS and others); wrappers for OSWorld, WindowsAgentArena, WebArena, AndroidWorld and more; Lite.OSWorld runs in a Docker desktop with no nested virtualization; datasets on Hugging Face; 126 stars |
| Runtime / SDK | Runtime / SDK | Free, Apache-2.0 (with listed third-party components), self-hostedPricing page | — | AutoPause suspends idle sandboxes and AutoResume wakes them on the next requestAutoPause suspends idle sandboxes and AutoResume wakes them on the next request; CubeCoW snapshots, clone and rollback; cross-node pause/resume with an S3 or JuiceFS backend (preview) | VM | Apr 2026 | <60 ms create from pooled snapshots, <5 MB overhead per sandbox; E2B SDK compatible (switch with an env var)<60 ms create from pooled snapshots, <5 MB overhead per sandbox; E2B SDK compatible (switch with an env var); eBPF egress policy and L7 credential injection; single node to multi-node cluster, Terraform deployer, ARM64; scheduler plugins; Prometheus/Grafana observability; 13,244 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); a managed EU service and an open-source Operator are on aFree, open source (MIT); a managed EU service and an open-source Operator are on a waitlist with no published price (price per hour for 2 vCPU / 4 GiB: n/a (runs on your own machine)) | N/a (runs on your own machine) | No managed lifecycle: the CLI starts, stops and removes local computers (start, stop, rmNo managed lifecycle: the CLI starts, stops and removes local computers (start, stop, rm, destroy); stateful machines are on the roadmap | Container | Jan 2025 | Framework for computer-use agents: TypeScript SDK (@cusedev/core), CLI (@cusedev/cli) and a REST API for displayFramework for computer-use agents: TypeScript SDK (@cusedev/core), CLI (@cusedev/cli) and a REST API for display, mouse, keyboard, bash and a file editor; keychain service built on the pass CLI so the agent can log in without credentials in the model context; noVNC and VNC through a local proxy on localhost:4242; Linux (Ubuntu) template only; built at a hackathon (Show HN 2025-01-19); 88 GitHub stars; last push 2025-03-03 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2026 | An open platform to run your agents densely on Kubernetes. (75 GitHub stars; Apache-2.0; owner dam-agents) |
| Runtime / SDK | Runtime / SDK | Free to use, including commercially, under the Deltavix Global Free-to-Use ProprietaryFree to use, including commercially, under the Deltavix Global Free-to-Use Proprietary License (no redistribution, no white-labeling, no SaaS resale); runs on your own Docker host; enterprise customization by contact | N/a (runs on your own servers) | Nodes are spun up and torn down through the Swarm Manager; no idle policy stated | Container | Aug 2026 | Self-hosted desktops for computer-use agents: Delta Control Protocol API per node with an optimized AT-SPI2Self-hosted desktops for computer-use agents: Delta Control Protocol API per node with an optimized AT-SPI2 accessibility tree, screenshots, Bezier-curve humanized mouse and typing, CAPTCHA "hesitation" routine and shell commands; MCP over SSE; per-node sticky or rotating proxies; web Cloud Console with live view and manual takeover; about 400 MB RAM per node; Docker on Linux, macOS or WSL2; prebuilt image deltavixglobal/deltabox (no source code in the repo); GitHub 5 stars, last push 2026-08-10 |
| Runtime / SDK | Runtime / SDK | Open source (Apache-2.0 OR MIT badge); you pay AWS for the EC2 pool | — | Warm pool of ready EC2 instances; each devbox is used once and terminated on releaseWarm pool of ready EC2 instances; each devbox is used once and terminated on release, and the reconciler launches a replacement | VM | Jun 2026 | Rust control plane (HTTP API and pool reconciler) and CLI; claim latency under 1 sRust control plane (HTTP API and pool reconciler) and CLI; claim latency under 1 s; SSH with short-lived Vouch certificates; tools baked into EBS snapshots; Terraform for AWS in smoketurner/devbox-infra; inspired by Stripe's Minions; repo homepage devbox.farm does not resolve; 1 GitHub star |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hosted; ships as the code-execution service of DifyPricing page | N/a (self-hosted) | Stateless: each request writes the code to a temporary file and runs it onceStateless: each request writes the code to a temporary file and runs it once; worker_timeout 5 s, max_workers 4 and max_requests 50 in the sample config | Process | Jan 2024 | HTTP service (port 8194, API key) that runs untrusted Python 3 and Node.js code for multi-tenant useHTTP service (port 8194, API key) that runs untrusted Python 3 and Node.js code for multi-tenant use; powers the Code node in Dify workflows; network on by default, optional SOCKS5/HTTP proxy; extra syscalls via ALLOWED_SYSCALLS; Linux only; Dify's docs warn that sandboxed code can still reach other services in the Dify deployment unless its network is isolated; latest release 0.2.15 (2026-04-27); GitHub 1,283 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hosted on your own Linux x86_64 machinesPricing page | N/a (self-hosted) | Sandboxes are permanent and cool down one step at a time: frozen after 10 minutes idleSandboxes are permanent and cool down one step at a time: frozen after 10 minutes idle (processes suspended, memory pushed to swap: 1 GiB resident drops to ~5 MiB, wake in ~50 ms), stopped after 3 days, archived to S3-compatible storage (tar + zstd) 7 days after stopping if a bucket is configured; acquireSandbox(key) brings back any state; only destroy loses data; all three timers set per sandbox; sandboxes made through the E2B API get E2B-style deadlines, with onTimeout pause supported | gVisor | Jul 2026 | One daemon with a SQLite ledger and one port, installed with one commandOne daemon with a SQLite ledger and one port, installed with one command; E2B-compatible API: the official e2b SDK works by changing two URLs (commands, PTY, files, watch, signed upload/download URLs, port proxy with wake-on-traffic, metrics); native TypeScript SDK (@dormice/sdk, not on npm yet), dor CLI and web console; several machines behind one gateway, each sandbox staying on the machine that made it; dor doctor checks the install; agent skill; README says early development, not production-ready; 1,903 GitHub stars; the repo's docs site is set up for dormice.dev, which did not resolve on 2026-11-14 |
| Runtime / SDK | Runtime / SDK | No public pricing; commercial product sold through sales (book a meeting)No public pricing; commercial product sold through sales (book a meeting); runs on your own Kubernetes nodes, so infrastructure is billed by your cloud | — | Edera runs under the upstream kubernetes-sigs Agent Sandbox controllerEdera runs under the upstream kubernetes-sigs Agent Sandbox controller, which provides pause/resume of Sandbox pods; Edera's own pages describe no suspend or snapshot feature; memory ballooning guide exists | VM | — | Kubernetes RuntimeClass edera routes pods into zones; agent-sandbox guide installs kubernetes-sigs Agent SandboxKubernetes RuntimeClass edera routes pods into zones; agent-sandbox guide installs kubernetes-sigs Agent Sandbox (v0.5.0rc1 pinned) on an Edera cluster; install guides for AWS EKS, Google Compute Engine, Azure Linux, Linode and other managed Kubernetes; GPU multi-tenancy and confidential computing use cases; OCAP 'bounded authority' credential approach for agents; open-source pieces on GitHub (krata Xen control plane GPL-2.0, styrolite, am-i-isolated, sprout); Gartner Cool Vendor 2026 (site claim) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Dec 2025 | A Python sandbox using Wasmtime (85 GitHub stars; Apache-2.0; owner eryx-org) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | High-performance sandbox runtime on Kubernetes with microVMs. (66 GitHub stars; Apache-2.0; owner opensandbox-group) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-hosted with Docker ComposeFree, open source (MIT); self-hosted with Docker Compose; uses your own Claude credentials or subscription | N/a (runs on your own servers) | Long-running workers; desktops attach and detach; no idle policy stated | Container | Feb 2026 | Orchestrates fleets of Claude Code and Claude computer-use agents (claude-agent-sdk) across containerized LinuxOrchestrates fleets of Claude Code and Claude computer-use agents (claude-agent-sdk) across containerized Linux desktops and VNC-reachable machines; live desktop streaming to a dashboard through Apache Guacamole; supervisor agent delegates tasks; NATS JetStream messaging; PostgreSQL; scheduled tasks, self-healing and self-learning; Telegram gateway for human-in-the-loop; patchright-cli browser automation; README warns not to expose it to untrusted networks; GitHub 155 stars; last push 2026-03-11 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | The sandbox agent framework. (8667 GitHub stars; Apache-2.0; owner withastro) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Nov 2023 | MicroVM pods for agent sandboxes and containers as a service: FeOS, the node layer in Rust, and FeRunMicroVM pods for agent sandboxes and containers as a service: FeOS, the node layer in Rust, and FeRun, its control plane (25 GitHub stars; Apache-2.0; owner gardenlinux) |
| Runtime / SDK | Runtime / SDK | Free to self-host under the Functional Source License 1.1 with an Apache-2.0 futureFree to self-host under the Functional Source License 1.1 with an Apache-2.0 future license (FSL-1.1-ALv2): internal and commercial use of your own agents allowed, competing hosted offerings prohibited | N/a (runs on your own machine or server) | No hosted lifecycle: the container runs until you stop itNo hosted lifecycle: the container runs until you stop it; without a named volume nothing persists, and a named volume keeps /home/agent (browser profile, cookies, downloads, settings) across restarts | Container | Mar 2026 | Rust MCP server that lets any LLM agent see the screen, move the mouse, type, launch apps and run shell commandsRust MCP server that lets any LLM agent see the screen, move the mouse, type, launch apps and run shell commands; one Docker command; noVNC view in the browser (interactive, not read-only); dev and secured postures with auth and TLS; works with cloud models (Claude, GPT, Gemini) or local models (llama.cpp); SpecterChat client can drive it; built on YV17Labs' NestRS framework; GitHub 158 stars; last push 2026-09-23 |
| Runtime / SDK | Runtime / SDK | Free, open source (README says MIT; the repo has no LICENSE file); self-hostedFree, open source (README says MIT; the repo has no LICENSE file); self-hosted; you pay your own cloud | N/a (self-hosted) | Sandboxes can be paused, resumed, stopped and restarted through the APISandboxes can be paused, resumed, stopped and restarted through the API; snapshots freeze a sandbox and restore it later; no automatic idle timeout documented | VM | Mar 2026 | Self-hosted alternative to E2B: 'Give your AI agents a full Linux VM'; ~125ms microVM start, ~1s DockerSelf-hosted alternative to E2B: 'Give your AI agents a full Linux VM'; ~125ms microVM start, ~1s Docker; control plane (gitvmd) auto-scales nodes on AWS, GCP or Azure; REST API for exec, files and snapshots; Go SDK only; custom rootfs templates; dashboard; from the GitAgent org (maintained by Lyzr); 3 GitHub stars, 9 commits, last push 2026-03-10 |
| Runtime / SDK | Runtime / SDK | Free, open source (README says MIT; GitHub detects no license file)Pricing page | N/a (self-hosted) | Default sandbox lifetime 3600 s (sandbox_timeout_seconds), adjustable through a TTL APIDefault sandbox lifetime 3600 s (sandbox_timeout_seconds), adjustable through a TTL API, -1 = never expires; persistent-mode metadata in Redis survives API restarts; single execution timeout 30 s | Container | Mar 2026 | REST API for one-shot and persistent sandboxes with sync or SSE outputREST API for one-shot and persistent sandboxes with sync or SSE output; Python, Node.js/TypeScript and Bash in one container; pre-warmed container pool; network modes isolated, open, whitelist, block-private; workspace sync to Local/S3/COS/OBS/OSS/MinIO; Helm chart with HPA; Chinese-language README; 1 star |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Agent Executor (AX) is Google's open agentic orchestration runtime (14705 GitHub stars; Apache-2.0; owner google) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); you pay Google Cloud and your model providerPricing page | N/a (billed by Google Cloud) | Warm sandbox pools keep a ready sandbox for low-latency startsWarm sandbox pools keep a ready sandbox for low-latency starts; multi-turn sessions checkpoint and resume, and another process can re-attach to a session or adopt a running turn | — | Sep 2026 | Python library (3.12+) with one Engine, Session and Run API for running Claude Code or Codex locally or deployed to GCPPython library (3.12+) with one Engine, Session and Run API for running Claude Code or Codex locally or deployed to GCP Agent Sandbox; Claude via Vertex AI by default, OpenAI and OpenRouter models; per-run secrets; skills, MCP servers, repos and packages declared on the spec; streaming, structured output, steering and exec(); setup tool harness-run-gcp-setup; notes that Agent Sandbox is not yet GA; 6 GitHub stars; v0.4.0 on PyPI |
| Runtime / SDK | Runtime / SDK | Free, self-hosted; the repo has no license filePricing page | N/a (self-hosted; idle VMs are snapshotted to your own S3 storage) | An idle monitor snapshots VMs that have proxy routes once they pass HATCH_IDLE_TIMEOUTAn idle monitor snapshots VMs that have proxy routes once they pass HATCH_IDLE_TIMEOUT (45 min by default); the snapshot (CPU registers, memory and the writable overlay disk) goes to S3-compatible storage and microVM is stopped; the next HTTP request or SSH connection restores the VM with the same IP, MAC and SSH port; VMs with active SSH sessions are skipped | VM | Feb 2026 | Go daemon (hatchd) wrapping microVM with a REST API (create, stop, snapshot, restore, routes)Go daemon (hatchd) wrapping microVM with a REST API (create, stop, snapshot, restore, routes); needs a Linux host with KVM, Postgres and optionally S3/MinIO; cloud-init users and SSH keys; per-VM SSH port forwarding with wake-on-SSH; subdomain reverse proxy with wake-on-request; TypeScript SDK; default VM 1 vCPU / 512 MiB; roadmap lists API keys and dashboard, multi-node scheduling, Cloud Hypervisor and OCI images; 19 GitHub stars; hatchvm.com is down (apex does not connect, www returns Cloudflare error 525) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | microVM orchestrators to boot from docker images, packed with L7 reverse proxy (43 GitHub stars; Apache-2.0microVM orchestrators to boot from docker images, packed with L7 reverse proxy (43 GitHub stars; Apache-2.0; owner herd-core) |
| Runtime / SDK | Runtime / SDK | Free, open source (BSD-3-Clause)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Lightweight Linux sandbox for AI agents. Kernel-native isolation (namespaces, cgroups, seccompLightweight Linux sandbox for AI agents. Kernel-native isolation (namespaces, cgroups, seccomp, Landlock) with REST API, MCP bridge, and web dashboard. Single Rust binary. (61 GitHub stars; BSD-3-Clause; owner TetiAI) |
| Runtime / SDK | Runtime / SDK | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (Apache-2.0 per README); runs locally or in your own cloud with no Hiver API key | — | Snapshots: the working tree is saved to a tarball on shutdown and restored on start (forSnapshots: the working tree is saved to a tarball on shutdown and restored on start (for example to GCS); homepage says Hiver VMs 'resume instantly'; no idle timeout stated | Container | May 2026 | Open-source agent runtime that records a replayable trace of model calls, subagents, browser actions, filesOpen-source agent runtime that records a replayable trace of model calls, subagents, browser actions, files, network requests and tool calls; inspector UI with live terminal; built-in images for Claude Code, Codex, Copilot CLI, OpenClaw, Chromium with CDP, Node and Python; egress policy with credentials injected by a proxy; FUSE volumes with ACLs; TypeScript and Python client SDKs; 1 star |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); runs on your own Mac, Linux box, cloud VM or bare metalPricing page | N/a (self-hosted) | Idle sandboxes auto-pause after a configurable sleepAfter and resume on the nextIdle sandboxes auto-pause after a configurable sleepAfter and resume on the next operation with the workspace intact; they never self-destroy (destroy is explicit); stop/start keeps /workspace on a named volume that survives daemon restarts; backups snapshot /workspace to host tarballs | Container | Jun 2026 | One daemon runs as many sandboxes as the machine holds and refuses new ones when memory is fullOne daemon runs as many sandboxes as the machine holds and refuses new ones when memory is full; egress gateway gives each sandbox a capped, expiring, revocable token instead of your real model or GitHub key; optional default-deny egress on Linux; REST API, CLI, TypeScript (npm) and Python SDKs; exec with sessions, file ops and watch, background processes, preview URLs, stateful Python/JS code interpreter; per-sandbox CPU/memory/PID limits; web dashboard with live terminal and cost meter; per-sandbox audit log to JSONL or OpenTelemetry; hotcell create --opencode; sandbox layer for Directionally Local's Autopilot coding agent; v0.1.40; 19 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2025 | A multi-hypervisor VM runtime for OCI images, supporting Cloud Hypervisor, microVM, QEMUA multi-hypervisor VM runtime for OCI images, supporting Cloud Hypervisor, microVM, QEMU, and Apple Virtualization.framework. (376 GitHub stars; MIT; owner kernel) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs on your own Ubuntu/Debian hosts with KVMPricing page | N/a (self-hosted) | No pause or idle model stated; forks are thin copy-on-write disk layers scheduled nearNo pause or idle model stated; forks are thin copy-on-write disk layers scheduled near the source VM; the README says Hyper does not support RAM snapshots | VM | Jun 2026 | Distributed microVM orchestrator written in Elixir: nodes that join the cluster become VM runnersDistributed microVM orchestrator written in Elixir: nodes that join the cluster become VM runners; affinity-based scheduling; OCI image loading; OpenTelemetry traces; only Postgres as an external dependency; gRPC interface (unauthenticated, off by default); the README places it in the same niche as Daytona and Runloop and argues against VM size limits (e.g. 128-core VMs); Hex package hypervm 0.1.0 (2026-07-09); 174 GitHub stars; developed by harmont.dev, whose domain does not resolve |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | A secure* runtime for autonomous AI agents. Policy from plain-English constitutionsA secure* runtime for autonomous AI agents. Policy from plain-English constitutions. (*https://ironcurtain.dev) (663 GitHub stars; Apache-2.0; owner provos) |
| Runtime / SDK | Runtime / SDK | Free, open source (README says Apache-2.0Free, open source (README says Apache-2.0; GitHub shows no detected license), self-hosted single Go binary; a public demo at isobox.hsingh.app | N/a (self-hosted) | One-shot executions are ephemeral with a wall-time cap (default 10 s, max 20 s per call)One-shot executions are ephemeral with a wall-time cap (default 10 s, max 20 s per call); stateful sessions keep a persistent /workspace across steps, live-kernel sessions keep variables, and re-attachable volumes plus a KV store persist data; no timeouts for sessions stated | gVisor | Jun 2026 | Self-hostable code-execution sandbox for AI agents: POST /execute with SSE streaming, /v1/sessions, /v1/memorySelf-hostable code-execution sandbox for AI agents: POST /execute with SSE streaming, /v1/sessions, /v1/memory, /v1/volumes; Python (batteries-included), JavaScript, TypeScript, Ruby, Go, Rust; per-call limits up to 512 MiB, 2 CPUs, 256 pids; opt-in firewalled egress to the public internet only; warm pool and Valkey Streams queue (default off); Prometheus metrics; OpenAPI 3.1 and llms.txt; dependency-free Python and JS clients; threat model and adversarial audit docs; 1 GitHub star; single day of commits (created and last pushed 2026-06-04) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0), self-hosted on your own Kubernetes cluster (Helm chart)Free, open source (Apache-2.0), self-hosted on your own Kubernetes cluster (Helm chart); no hosted service | N/a (self-hosted) | Sandboxes are deleted explicitly or when timeout_seconds expiresSandboxes are deleted explicitly or when timeout_seconds expires; a SnapshotRootfs termination policy can snapshot the root filesystem to object storage on teardown, and rootfs snapshots restore into new sandboxes on any node (pre-warm an environment or resume a working session); no pause that keeps memory | gVisor | Nov 2025 | Self-hosted Kubernetes sandboxing for untrusted and AI-generated code; any OCI imageSelf-hosted Kubernetes sandboxing for untrusted and AI-generated code; any OCI image; configurable CPU, memory and ephemeral storage; multi-container sandboxes (app plus sidecars and MCP servers); REST API for commands, files and output streams; Python (isola-run) and TypeScript (@isola-run/sdk) SDKs; sandboxes and snapshots are Kubernetes custom resources run by an operator; no database or queue needed; 61 GitHub stars; latest release v0.8.1 (2026-11-06) |
| Runtime / SDK | Runtime / SDK | Free, open source (npm package says Apache-2.0Free, open source (npm package says Apache-2.0; the repo has no LICENSE file at its root); runs inside your own Node.js process | N/a (self-hosted library) | A Bash instance lives in your process; the filesystem persists across exec() calls whileA Bash instance lives in your process; the filesystem persists across exec() calls while env, functions and cwd reset per call; every invocation bounded by maxExecutionTimeMs and other execution limits | Process | Dec 2025 | Vercel Labs project; npm install just-bash (v3.6.0); 'Designed for AI agents needing a secure, sandboxed environment'Vercel Labs project; npm install just-bash (v3.6.0); 'Designed for AI agents needing a secure, sandboxed environment'; unix commands (grep, awk, jq, sed, rg and more) plus optional curl, Python, JS/TS and sqlite; network off by default, URL-prefix and method allow-lists when on; OverlayFs copy-on-write over a real directory; custom commands; Vercel Sandbox-compatible API; bash-tool wraps it for the AI SDK; Node >= 20.19; README points to Vercel Sandbox for a full VM; GitHub vercel-labs/just-bash 4,483 stars |
| Runtime / SDK | Runtime / SDK | Free, Apache-2.0, self-hostedPricing page | — | k7d backend: pause/resume in place keeps memory (0.2 s / 0.3 s)k7d backend: pause/resume in place keeps memory (0.2 s / 0.3 s); QEMU+Longhorn backend: pause 1.3 s, disk survives; Kata-microVM backend: scale to 0/1 | VM | Oct 2025 | Warm VM fork (copy-on-write disk and memory, ~5 ms at VMM, ~2.4 s end to end) on k7d; Docker inside the VMWarm VM fork (copy-on-write disk and memory, ~5 ms at VMM, ~2.4 s end to end) on k7d; Docker inside the VM; named snapshots and restore on QEMU+Longhorn path; CLI, API, Python SDK (k7-sdk); Show HN #1 Oct 2025; 826 GitHub stars; docs at docs.katakate.org |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2026 | Rootless container runtime with no daemon. Runs agent tool calls and LLM-generated code in kernel-enforced containersRootless container runtime with no daemon. Runs agent tool calls and LLM-generated code in kernel-enforced containers. (484 GitHub stars; Apache-2.0; owner getkern) |
| Runtime / SDK | Runtime / SDK | Free, open source: KernelBox MIT (PyPI metadata and README badgeFree, open source: KernelBox MIT (PyPI metadata and README badge; no license file in the repo), agentnb MIT, snipara-sandbox Apache-2.0 | N/a (self-hosted library) | All three keep a Python session alive between calls: KernelBox named kernels untilAll three keep a Python session alive between calls: KernelBox named kernels until destroy(); agentnb project-scoped sessions with a 30 s default execution timeout; snipara-sandbox persistent, resumable workflow state | Container | Jan 2026 | Three unrelated small Python packages for agents grouped in one row. KernelBox 0.1.2 (Vinay ChaudhariThree unrelated small Python packages for agents grouped in one row. KernelBox 0.1.2 (Vinay Chaudhari; FastAPI, CLI and Python API; 3 stars; last release 2026-05-17). agentnb 0.3.10 (Oege Dijk; CLI-driven persistent REPL for coding agents; 7 stars; last release 2026-03-23). snipara-sandbox 3.1.0 (Snipara, a project-memory product for coding agents; needs local Docker; LiteLLM inference optional; 2 stars; last release 2026-11-13) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); you pay for your own Docker host or Modal/E2B accountsPricing page | N/a (self-hosted) | Ephemeral: the sandbox lives for the MCP session and is deleted when the connectionEphemeral: the sandbox lives for the MCP session and is deleted when the connection closes; idle session timeout 300 s (HTTP mode); E2B sandbox lifetime default 3600 s | Container | Feb 2026 | MCP server with one sandbox_exec tool; one sandbox per agent connectionMCP server with one sandbox_exec tool; one sandbox per agent connection; agent talks to the sandbox over MCP instead of running inside it, keeping keys and prompts out; stdio and streamable HTTP; any OCI image; GPU via Modal backend; Python 3.13+, PyPI; MIT; 52 stars; last push 2026-03-03 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs on your own machinesPricing page | N/a (self-hosted) | Ephemeral: each Bash tool call mounts a Workspace, runsEphemeral: each Bash tool call mounts a Workspace, runs, commits its changes and unmounts; history is kept as Commits in a SQLite Store (Disposable profile, no crash or power-loss durability) | Container | Aug 2026 | Layered copy-on-write filesystem for agents; every tool call becomes a Commit on a Branch; branch, rewind and mergeLayered copy-on-write filesystem for agents; every tool call becomes a Commit on a Branch; branch, rewind and merge; content-deduplicated store (157 snapshots in 100 MB, 19.9x smaller than Btrfs + zstd, own benchmark); runs today through the dsh-layerfs plugin for DeepSeek Harness; Linux, macOS through DSH, Windows through WSL2; JavaScript SDK (@layerfs/agent, @layerfs/backend, @layerfs/runtime); v1.1.0; MIT; 2,779 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2026 | Core Rust library for Lens sandbox policy enforcement, networking, DNS, proxying, and boundary credential exchangeCore Rust library for Lens sandbox policy enforcement, networking, DNS, proxying, and boundary credential exchange. (8 GitHub stars; Apache-2.0; owner lensapp) |
LibreChat Code Interpreter APIHosted API closed to new subscriptions Runtime / SDK | Runtime / SDK | Hosted API: 'Subscriptions Currently Unavailable', no new subscriptionsHosted API: 'Subscriptions Currently Unavailable', no new subscriptions, existing subscribers keep theirs; the self-hosted service is free (Apache-2.0) | — | Stateless by default: each run is a fresh processStateless by default: each run is a fresh process; experimental stateful sessions keep a /mnt/data workspace between runs (per user, per agent and user, or per conversation), which can be reset at any time | VM | Dec 2024 | Runs Python, Node.js (JavaScript and TypeScript), Go, C/C++, Java, PHP, Rust, Fortran and R; file upload and downloadRuns Python, Node.js (JavaScript and TypeScript), Go, C/C++, Java, PHP, Rust, Fortran and R; file upload and download; programmatic tool calling from inside the sandbox; egress gateway; Docker Compose and Helm chart; LibreChat is joining ClickHouse; open source as ClickHouse/code-interpreter (Apache-2.0, 162 GitHub stars) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT per the site); runs in the user's browser or your ownFree, open source (MIT per the site); runs in the user's browser or your own Node/serverless process, $0 infrastructure | N/a (runs locally) | Runs in a browser tab or Node process; virtual filesystem persisted to IndexedDB soRuns in a browser tab or Node process; virtual filesystem persisted to IndexedDB so sessions survive reloads; no hosted lifecycle | Process | Feb 2026 | Boots in ~0.5 ms in Node and ~27 ms in a browser tab, 0.3 MB gzipped (own benchmarks)Boots in ~0.5 ms in Node and ~27 ms in a browser tab, 0.3 MB gzipped (own benchmarks); Bash-like shell, 60+ Unix commands including git and curl; Node.js compatibility layer; real npm and npx from the registry; live previews of in-VM ports through a service worker; runs Vite, Expo and Supabase (tinbase) unmodified; cannot run native Linux binaries; Sandbox API in @lifo-sh/core (TypeScript); single-binary and `npx lifo-sh` modes; alpha (v0.10); 564 GitHub stars; by Sanket Sahu (RapidNative) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | N/a (self-hosted) | Session-scoped: the sandbox lives as long as the SandboxSession contextSession-scoped: the sandbox lives as long as the SandboxSession context; container pooling pre-warms and reuses containers; execution time limits configurable | Container | Jun 2024 | Python library; runs Python, JavaScript, Java, C++, Go and R with dependency install; MCP serverPython library; runs Python, JavaScript, Java, C++, Go and R with dependency install; MCP server; interactive IPython sessions (Docker, Podman, Kubernetes); plot/artifact extraction; security policies, CPU, memory and time limits, network isolation; remote Docker host; examples for 11 agent frameworks (OpenAI Agents SDK, Claude Agent SDK, LangChain, LlamaIndex, CrewAI and others); MIT; 1,144 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Lightweight AgentFS sandbox that runs bash and python. (157 GitHub stars; no license file; owner coplane) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | Matchlock secures AI agent workloads with a Linux-based sandbox. (652 GitHub stars; MIT; owner jingkaihe) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); you pay for your own Docker host or Fly/Modal/E2B accounts (FlyFree, open source (MIT); you pay for your own Docker host or Fly/Modal/E2B accounts (Fly default 1 shared CPU/512 MB) | N/a (self-hosted) | Temporary computers are deleted when the MCP session closesTemporary computers are deleted when the MCP session closes; permanent computers survive and can be reattached by ID; restart keeps writable state; factory reset recreates from the image; idle session timeout 300 s (HTTP mode); sandbox lifetime timeout default 3600 s, max 86400 s | Container | Aug 2026 | Fork of Kilntainers; named computers with stable IDs; MCP App dashboard (list, run, restart, factory reset, delete)Fork of Kilntainers; named computers with stable IDs; MCP App dashboard (list, run, restart, factory reset, delete); terminal_execute tool plus optional computer_* lifecycle tools; Fly Machines backend with persist_rootfs for permanent machines; PyPI; MIT; 3 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | N/a (self-hosted) | Run-to-completion sandboxes killed at their timeout or TTL (outcome states includeRun-to-completion sandboxes killed at their timeout or TTL (outcome states include timed_out); no pause or suspend described | VM | Jul 2026 | Go daemon and CLI; REST + SSE API generated from one OpenAPI spec; Go, Python and TypeScript SDKsGo daemon and CLI; REST + SSE API generated from one OpenAPI spec; Go, Python and TypeScript SDKs; Go/Python/TypeScript/Rust images; README benchmarks: boot 170-300 ms, about 214 ms per create end to end; HTTP over vsock control channel; pull-based Redis queue scales from one VPS to many nodes with CPU/memory packing; active-CPU metering; 1 star |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2025 | A modular and stable agent sandbox runtime environment. (58 GitHub stars; Apache-2.0; owner modelscope) |
| Runtime / SDK | Runtime / SDK | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2026 | Isolated short-lived agent sandboxes as pods in your own Kubernetes cluster (mogenius) [from repo/listing] (1 GitHubIsolated short-lived agent sandboxes as pods in your own Kubernetes cluster (mogenius) [from repo/listing] (1 GitHub stars; NOASSERTION; owner mogenius) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2026 | NEAR AI Red (near-ai-red): microVM sandbox for agentic security research and red teaming. (12 GitHub stars; Apache-2.0NEAR AI Red (near-ai-red): microVM sandbox for agentic security research and red teaming. (12 GitHub stars; Apache-2.0; owner nearai) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); you run it on your own hosts (the supported path is oneFree, open source (Apache-2.0); you run it on your own hosts (the supported path is one Latitude.sh bare-metal host provisioned from a signed release) | N/a (self-hosted) | Machines self-destruct on a TTL by defaultMachines self-destruct on a TTL by default; 'keep alive' runs one until it is stopped when the server sets NEHEMIAH_ALLOW_PERSISTENT; machines restore from a memory snapshot in about 3 ms; a running machine forks with its live state in about 35 ms; S3-backed persistent volumes outlive a machine | VM | Jun 2026 | On-demand Linux computers for AI: a full desktop over VNC with a browser, or a headless shellOn-demand Linux computers for AI: a full desktop over VNC with a browser, or a headless shell; claude, codex, cursor and pi preinstalled; a built-in AI that drives the machine and returns a live URL; files and ports through the daemon; Go host daemon (nehemiahd), SvelteKit site, Effect TypeScript SDK (nehemiah-sdk) and MCP server (nehemiah-mcp); 347 GitHub stars, 48 forks; last push 2026-09-19; the earlier one-command self-serve installers are no longer supported; boringcomputers.com, which the README calls a showcase, now shows only 'AI consultancy and software factory', and its /docs returns 404 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Run agents like Hermes, LangChain Deep Agents, and OpenClaw more securely inside NVIDIA OpenShell with managedRun agents like Hermes, LangChain Deep Agents, and OpenClaw more securely inside NVIDIA OpenShell with managed inference (23027 GitHub stars; Apache-2.0; owner NVIDIA) |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0); self-hosted with Docker Compose or HelmFree, open source (AGPL-3.0); self-hosted with Docker Compose or Helm; no hosted service (the billing docs still point at Daytona's dashboard) | N/a (self-hosted) | Inherited from Daytona: auto-stop after 15 min without user interaction by default (0Inherited from Daytona: auto-stop after 15 min without user interaction by default (0 disables), auto-archive after 7 days stopped by default (max 30 days), auto-delete off by default; ephemeral sandboxes are deleted on stop | Container | Jul 2026 | Community fork of Daytona's last AGPL release (v0.190.0) after Daytona moved development private in June 2026Community fork of Daytona's last AGPL release (v0.190.0) after Daytona moved development private in June 2026; keeps the full Daytona history; SDKs in Python, TypeScript, Ruby, Go and Java (still published under Daytona names); REST API, CLI, dashboard, web terminal, SSH, VNC, preview URLs, snapshots, volumes, computer use, MCP server, LSP; default sandbox 1 vCPU/1GB RAM/3GiB disk; under 90ms start claimed; not affiliated with Daytona; 27 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); runs on your own LinuxFree, open source (Apache-2.0); runs on your own Linux, macOS (Apple silicon) or Windows/WSL2 machine, Kubernetes cluster, or a Brev Launchable in NVIDIA's cloud | — | Sandboxes are created, stopped, started and deleted explicitlySandboxes are created, stopped, started and deleted explicitly; stop keeps the sandbox record and storage; delete stops processes, releases resources and purges injected credentials; --no-keep deletes the sandbox when its command exits; no idle auto-suspend described | Container | Feb 2026 | Declarative per-agent policy for files, network and processesDeclarative per-agent policy for files, network and processes, with formal verification flagging risky policy changes for human review; agents never see real credentials (gateway injects them only for approved endpoints, including inference); GPU passthrough via CDI; SSH into sandboxes; Python, TypeScript, Go and Rust SDKs; CLI and agent skills; Helm chart for Kubernetes; 21,630 GitHub stars; moving to CNCF (ai-openshell org) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); no hosted servicePricing page | — | — | VM | Aug 2026 | Ultra-light microVM sandbox for untrusted agentic workloads from MSR Systems Research and Azure Research SystemsUltra-light microVM sandbox for untrusted agentic workloads from MSR Systems Research and Azure Research Systems, building on Nanvix; guests Alpine, Ubuntu and Azure Linux; Rust CLI nvx-rs; runs on Linux and Windows hosts; only development pre-releases so far (latest 2026-11-14); 841 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-hosted (Docker image onyxdotapp/code-interpreter or Helm)Pricing page | N/a (self-hosted) | Each execution gets a fresh container that is removed afterwardsEach execution gets a fresh container that is removed afterwards; timeout per request (max 10 s by default), CPU time 5 s, memory 128 MB by default | Container | Sep 2025 | REST API (POST /v1/execute) for untrusted Python from LLMs, with file upload and downloadREST API (POST /v1/execute) for untrusted Python from LLMs, with file upload and download; common Python libraries preinstalled; Docker-out-of-Docker or Docker-in-Docker deployment; powers the Code Interpreter tool in Onyx; GitHub 30 stars |
| Runtime / SDK | Runtime / SDK | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (MIT); E2B usage billed by E2B | — | Microsandbox: idle sandboxes release compute and resume from checkpointsMicrosandbox: idle sandboxes release compute and resume from checkpoints, on a compatible node, with workspace files kept on NFS; E2B: paused in place by E2B; OpenSandbox: no suspend, a sandbox with a Session keeps running | VM | Sep 2026 | Open-source, self-hosted implementation of OpenAI's Agents API (official OpenAI SDK via OPENAI_BASE_URL)Open-source, self-hosted implementation of OpenAI's Agents API (official OpenAI SDK via OPENAI_BASE_URL); runs Codex, Claude Code or MiniMax Code harnesses with your own model provider; Core with PostgreSQL plus a Web console; pluggable Sandbox Provider, workspace filesystem, harness and model protocols; installers for Linux/macOS (Docker) and Windows; MIT; 326 stars; by MiniMax |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0), self-hosted on your own Kubernetes clusterFree, open source (Apache-2.0), self-hosted on your own Kubernetes cluster; no hosted service | N/a (self-hosted) | Pause and resume on demand (E2B pause()/connect()Pause and resume on demand (E2B pause()/connect(), spec.paused) or automatically after a sandbox reports idle via runtime probes, with scheduled wake-up and wake on inbound traffic through the Sandbox Gateway; pause strategies: Default persists the rootfs to the backing volume, Stop keeps only PVC data, Hibernate/Snapshot persists the rootfs to a snapshot; the site claims hibernation and checkpoints covering memory, read/write layer and GPU memory; timeout-based auto-pause | Container | Nov 2025 | CNCF OpenKruise sub-project for AI agent workloads on Kubernetes: warm pools (SandboxSet), sandbox claimCNCF OpenKruise sub-project for AI agent workloads on Kubernetes: warm pools (SandboxSet), sandbox claim, user identity and session management with traffic routing, E2B-compatible API (code interpreter and desktop examples) plus Kubernetes CRD API and Go SDK; agent-runtime sidecar for commands and filesystem; commit sandbox image, snapshots, storage management, in-place upgrades; user cases listed: RL training at Kimi, OpenClaw at Li Auto, MiniMax MaxClaw; GitHub openkruise/agents 319 stars; latest release v0.6.0 (2026-10-09) |
| Runtime / SDK | Runtime / SDK | Free, Apache-2.0, self-hostedPricing page | — | Sandbox timeout set at create (e.g. --timeout 30m)Sandbox timeout set at create (e.g. --timeout 30m); Fast Sandbox (microVM) pause checkpoints rootfs, VM state and memory to an artifact store and releases the node, resume on any node; snapshots | Container | Dec 2025 | microVM create 64 ms P50 serial, 125 ms P99 at 10 concurrent; commands, files, code interpreter, browsers and desktopsmicroVM create 64 ms P50 serial, 125 ms P99 at 10 concurrent; commands, files, code interpreter, browsers and desktops; ingress gateway and per-sandbox egress policies; credential vault; resource pools and batch creation for evals and RL; SDKs for Python, Java/Kotlin, TypeScript, C#/.NET, Go; osb CLI; MCP; packages still published under Alibaba names (com.alibaba.opensandbox, @alibaba-group/opensandbox); 17,302 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | A sandboxed Python runtime for AI agents, written in Rust. (146 GitHub stars; MIT; owner parcadei) |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Create, manage, and orchestrate stereOS AI agent sandboxes. (77 GitHub stars; AGPL-3.0; owner papercomputeco) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Nov 2025 | Pctx is the execution layer for agentic tool calls. It auto-converts agent tools and MCP servers into code that runs inPctx is the execution layer for agentic tool calls. It auto-converts agent tools and MCP servers into code that runs in secure sandboxes for token-efficient workflows. (287 GitHub stars; MIT; owner portofcontext) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2018 | A high performance general purpose code execution engine. (2843 GitHub stars; MIT; owner engineer-man) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2018 | ML/AI training/serving and agent sandbox operator and controller for Kubernetes (95 GitHub stars; Apache-2.0ML/AI training/serving and agent sandbox operator and controller for Kubernetes (95 GitHub stars; Apache-2.0; owner polyaxon) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2023 | A secure Python sandbox, written in Rust, for code written by AI. (8926 GitHub stars; MIT; owner pydantic) |
| Runtime / SDK | Runtime / SDK | No pricing published; self-hosted via install script (license not stated) | — | Sandboxes start from warm template snapshots; SDK kills the sandbox on block exitSandboxes start from warm template snapshots; SDK kills the sandbox on block exit; no idle policy stated | VM | — | Alpha; control plane plus host agent on Linux x86_64 with KVM, no KubernetesAlpha; control plane plus host agent on Linux x86_64 with KVM, no Kubernetes; Python SDK, REST and WebSocket API, CLI, operator dashboard; stateful code interpreter; headless Chromium over CDP; object-storage volumes; preview URLs; no telemetry; built by Utibeabasi Umanah |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs on your own machinePricing page | N/a (self-hosted) | VM runs while the Python Sandbox context is openVM runs while the Python Sandbox context is open; save() writes disk state to reload later, even on another machine; checkpoint()/revert() captures full VM state; optional auto-save on stop | VM | Apr 2026 | Async Python API (pip install quick-sandbox); prebuilt Ubuntu 24.04 and Alpine 3.23 imagesAsync Python API (pip install quick-sandbox); prebuilt Ubuntu 24.04 and Alpine 3.23 images, desktop images with Xfce4 and a browser, quicksand-cua image with Xvfb, noVNC and Playwright; screenshot, keyboard and mouse control; boot p50 0.37 s (alpine) to 0.91 s on an M3 Max; network-isolated by default; host directory mounts; port forwarding both ways; several Linux users per VM; x86_64 and ARM64 on macOS, Linux and Windows; no root or Docker needed; MIT; 411 GitHub stars; v0.13.0 (2026-10-14) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT OR Apache-2.0); pricing of the hosted quilt.sh service could notFree, open source (MIT OR Apache-2.0); pricing of the hosted quilt.sh service could not be checked because the site is down | N/a (self-hosted) | — | Container | Jun 2025 | Open-source runtime behind quilt.sh: gRPC daemon and CLI, SQLite state engine, ICC networking (bridge, veth, DNSOpen-source runtime behind quilt.sh: gRPC daemon and CLI, SQLite state engine, ICC networking (bridge, veth, DNS, firewall), OCI image workflows, volumes, metrics; cluster workflows in the separate quiltc; optional qgui browser-accessible desktop (Xvfb/VNC/websockify); by Aria Compute Company; 17 stars; last push 2026-02-25 |
| Runtime / SDK | Runtime / SDK | Rig is free (MIT). You pay E2B for running timeRig is free (MIT). You pay E2B for running time; README: 'about $0.46 an hour for a 4 CPU / 16 GB cloud desktop at E2B's published rates'; paused desktops cost nothing | $0 while paused (README); E2B bills running time | A desktop pauses after 15 minutes without a Rig command (RIG_IDLE_MIN)A desktop pauses after 15 minutes without a Rig command (RIG_IDLE_MIN), keeping memory and running processes, and wakes in about 2 s; paused boxes unused for 7 days are deleted automatically (RIG_AUTO_PRUNE_DAYS) | — | Sep 2026 | CLI (Bun) that runs on your laptop and gives each git branch an Ubuntu desktop on E2B with the dev serverCLI (Bun) that runs on your laptop and gives each git branch an Ubuntu desktop on E2B with the dev server, tests and a signed-in Chrome, while the agent stays local; copies browser logins (Chrome, Arc, Edge, Brave, Firefox, Safari) into a default desktop; 'rig mcp' gives Claude Code computer-use tools (screenshot, click, type); 'rig desktop' opens the screen in a browser tab; 'rig sync' sends uncommitted edits; saved desktops; preinstalled Node, Bun, Python, Playwright, Claude Code, Codex and cloud CLIs; works with self-hosted E2B; npm @shadowwalker2014/rig (v0.1.2); MIT; 7 stars; last push 2026-09-25; repo moved from ShadowWalker2014/rig |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); GitHub bills the codespaces to your own account at GitHub's ratesPricing page | GitHub bills codespace storage while a stopped sandbox is retainedGitHub bills codespace storage while a stopped sandbox is retained, 24 h by default | Run creates a codespace, runs one command and deletes it (--keep always or failure stopsRun creates a codespace, runs one command and deletes it (--keep always or failure stops it instead); defaults: --idle 10 min, --retention 24 h after a sandbox stops, --timeout 2 h per command | VM | Oct 2026 | Single static Go binary with a CLI and an MCP server; turns disposable GitHub codespaces into sandboxes: createSingle static Go binary with a CLI and an MCP server; turns disposable GitHub codespaces into sandboxes: create, exec over SSH through the codespace's Dev Tunnel, cp files, diff returns a patch since create, delete; needs only a GitHub token with the codespace scope; default machine basicLinux32gb (2 cores); uses the undocumented codespace connection protocol, which can change without notice; MIT; 9 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2026 | Local-first, self-hosted AI agent runtime and MCP bridge with sandboxed sessions, memory, credentials, audit/replayLocal-first, self-hosted AI agent runtime and MCP bridge with sandboxed sessions, memory, credentials, audit/replay, and a local Console. (644 GitHub stars; Apache-2.0; owner sandbaseai) |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0); runs on your own Kubernetes cluster or a sandboxd meshPricing page | N/a (self-hosted) | Adds pause, resume, fork and snapshot subresources to the Kubernetes Sandbox APIAdds pause, resume, fork and snapshot subresources to the Kubernetes Sandbox API; sandboxes are claimed from node-local sandboxd warm pools; per-sandbox TTL and hibernation come from sandboxd (the Cocoon Sandbox row) | VM | Jul 2026 | Aggregated Kubernetes apiserver serving sandboxes.agents.x-k8s.io (the kubernetes-sigs/agent-sandbox API) from per-nodeAggregated Kubernetes apiserver serving sandboxes.agents.x-k8s.io (the kubernetes-sigs/agent-sandbox API) from per-node inventory with no per-sandbox etcd object; README claims one SandboxWarmPool patch took 20 bare-metal nodes to 50,000 running microVMs in 10-15 s with about 2 etcd writes/s; optional e2b-compatible REST surface plus sandbox-envd-proxy so the e2b SDK's files, commands and PTY traffic reach the sandbox; mesh mode without Kubernetes (sandbox-e2b); Helm chart; part of the cocoonstack microVM stack (cocoon engine, sandbox, vk-sandbox); 87 GitHub stars |
| Runtime / SDK | Runtime / SDK | Self-hosted open source; no hosted service or prices | — | — | Container | Jan 2025 | Runs AI-generated Python (IPython cells) and shell commands in Docker containersRuns AI-generated Python (IPython cells) and shell commands in Docker containers; Python client (sandboxai-client) and Go SDK; REST API with OpenAPI spec; Kubernetes backend listed as "coming soon"; no LICENSE file; 143 GitHub stars; no commits since 2025-02-05; substratus.ai did not load |
| Runtime / SDK | Runtime / SDK | Free, self-hosted (no license file detected by GitHub)Pricing page | N/a (self-hosted) | — | Container | Dec 2025 | Renamed Orb: sandboxed.sh is now the Rust backend ('Core') for the Orb macOS, iOS and Android clientsRenamed Orb: sandboxed.sh is now the Rust backend ('Core') for the Orb macOS, iOS and Android clients; orchestrates Claude Code, Codex, Antigravity, OpenCode and Grok on your computer or your servers, plus cloud agents (Hermes, Grok Bot, Cursor Cloud); remote nodes and mission transfer between machines; headless Wayland/Sway desktop for browser and GUI automation; REST API and MCP; subscriptions via CLIProxyAPI; 716 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hostedPricing page | N/a (self-hosted) | Stateless HTTP requests: each run has compile and run timeouts (10 s defaults) and anStateless HTTP requests: each run has compile and run timeouts (10 s defaults) and an optional memory limit | Container | Oct 2024 | ByteDance code sandbox for running and judging LLM-generated code in evaluation and RL trainingByteDance code sandbox for running and judging LLM-generated code in evaluation and RL training; about 20 languages (Python, C++, Go, Java, Rust, Verilog, CUDA and more) plus a Jupyter mode; built-in runners for 10+ coding benchmarks (HumanEval, MBPP, CodeContests and others) behind one HTTP API; online playground in the docs; described in arXiv 2412.00535; GitHub bytedance/SandboxFusion 1,083 stars, last push 2026-07-14 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-host only, no hosted servicePricing page | N/a (self-hosted) | Auto-suspend on idle (SANDKILN_AUTO_SUSPEND_TIMEOUT_SECS) pauses and snapshotsAuto-suspend on idle (SANDKILN_AUTO_SUSPEND_TIMEOUT_SECS) pauses and snapshots; stop auto-snapshots by default; optional idle-timeout destroy as a backstop; archive tier moves old snapshots to a separate directory | VM | Aug 2026 | Rust daemon sandkilnd; JS/TS SDK, Python SDK and CLI (kiln); snapshot, resume and fork with lineage; named sandboxesRust daemon sandkilnd; JS/TS SDK, Python SDK and CLI (kiln); snapshot, resume and fork with lineage; named sandboxes; read-only shared drives; custom images; per-sandbox I/O rate limits; egress firewall rules; cold create ~144 ms, warm-pool resume ~4-18 ms (changelog); needs KVM; one-developer project; MIT; 2 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); runs on your own Linux 6.12+ hostPricing page | N/a (self-hosted) | Runs a single process or pipeline under policyRuns a single process or pipeline under policy; checkpoint feature and copy-on-write fork; no hosted lifecycle | Process | Mar 2026 | ~5 ms startup claim; per-host and per-HTTP-method/path network rules with DNS pinned at start~5 ms startup claim; per-host and per-HTTP-method/path network rules with DNS pinned at start; copy-on-write workspace (seccomp backend or BranchFS FUSE); COW fork ~530 µs per fork claim; memory and resource limits; /proc virtualization; OCI runtime shim for containerd, CRI-O and Kubernetes; sandlock.mcp runs each MCP tool call in its own sandbox; Sandbox HTTP API for remote execution; Rust core and CLI, C FFI, Python and Go SDKs; 618 GitHub stars; arXiv paper 2605.26298 |
| Runtime / SDK | Runtime / SDK | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (Apache-2.0) | — | Four tiers: hot (running); frozen (paused, cold pages in zswap, ~0.1 s back)Four tiers: hot (running); frozen (paused, cold pages in zswap, ~0.1 s back); snapshot (checkpointed to disk, memory freed, ~0.2 s back); detached (only the volume remains, cold start). A gateway wakes a sandbox on incoming traffic | gVisor | May 2026 | Formerly forkd (github.com/deeplethe/forkd redirects here)Formerly forkd (github.com/deeplethe/forkd redirects here); one master (SQLite) and many nodes, nodes can dial out from behind NAT; single binary sandstormd with web console; warm templates; README benchmark: 84 ms warm start on a cloud VM; nodes refuse work when memory runs low; Apache-2.0; 6,095 stars; by DeepLethe |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0); self-hostedPricing page | N/a (self-hosted) | Idle desktops auto-pause after 20 minutes by default (idle_pause_minutes, 0 disables)Idle desktops auto-pause after 20 minutes by default (idle_pause_minutes, 0 disables); acquired desktops are released after a configurable TTL; snapshot save and restore of files and browser sessions | Container | Apr 2026 | MCP server over stdio, streamable HTTP or SSE with 21 tools and 100+ actions (screenshot, OCR, click, type, shellMCP server over stdio, streamable HTTP or SSE with 21 tools and 100+ actions (screenshot, OCR, click, type, shell, files, Chromium control, semantic page map); watch and take over via RDP or noVNC; multi-agent registration with API keys and desktop ownership; web dashboard; clone desktops from templates; knowledge compiled from past session logs; about 2 GB RAM per desktop, no GPU; Linux, macOS (Docker Desktop) and Windows (WSL2); pip screenbox-mcp; 26 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2026 | Sdkwork-sandbox (0 GitHub stars; NOASSERTION; owner sdkwork-ai) |
| Runtime / SDK | Runtime / SDK | Free, open source library (Apache-2.0, npm secure-exec)Free, open source library (Apache-2.0, npm secure-exec), runs on your own servers or wherever Node.js, Bun or a browser runs; no hosted price | N/a (self-hosted library) | Evaluate/execute run each call in a fresh VM that is disposed when the call finishesEvaluate/execute run each call in a fresh VM that is disposed when the call finishes (per-call timeoutMs; CPU time budgets and memory caps, runaway code exits with code 124); createVm keeps files, installed packages and running servers until you dispose it | V8 isolate | Feb 2026 | Library for running untrusted JavaScript and TypeScript from AI agents with real Node.js APIs (fs, child_process, httpLibrary for running untrusted JavaScript and TypeScript from AI agents with real Node.js APIs (fs, child_process, http, dns, process, os) and npm packages; host functions with zod schemas so credentials stay in the host process; works with Vercel AI SDK, LangChain and other tool-use frameworks; site claims 17.9 ms p50 cold start and about 3.4 MB per execution; requires Node.js 22+ on Linux or macOS; code now lives in rivet-dev/agentos (4,899 stars, shared with agentOS); the old rivet-dev/secure-exec repo now redirects to rivet-dev/dynamic-apps; secureexec.dev redirects to rivet.dev/secure-exec |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | N/a (in-process) | In-process Go value; per-call timeout and limits; no lifecycle beyond the host process | Process | Jul 2026 | Go library (Go 1.25+); deny-by-default commands, filesystem and networkGo library (Go 1.25+); deny-by-default commands, filesystem and network; python3 and js interpreters under the same confinement with a memory cap; host mounts read-only, read-write or overlay, confined with os.Root; network allow/deny rules with SSRF guard and header injection; SDK-neutral tool API for anthropic-sdk-go, openai-go and agent-framework-go; MIT; 44 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source demo; you pay your own cloud (the blog's example is an AWS m6i.xlargeFree, open source demo; you pay your own cloud (the blog's example is an AWS m6i.xlarge at $0.192/h, an AWS price, not a SkyPilot one) | N/a (self-hosted) | Sessions are pooled and kept warm between executionsSessions are pooled and kept warm between executions; the blog says it 'shuts down when idle' via auto-scaling | Container | Aug 2025 | Proof-of-concept from the SkyPilot team's blog: SkyPilot launches the sandbox on your own cloud or KubernetesProof-of-concept from the SkyPilot team's blog: SkyPilot launches the sandbox on your own cloud or Kubernetes, llm-sandbox runs the code, and an MCP server exposes it to Claude Desktop, VS Code and other MCP clients; Python, JavaScript, Java, C++, Go and R; read-only S3 mounts for team data; the blog lists limits, including sessions pooled across users; repo alex000kim/skypilot-code-sandbox, 17 stars, no license file, last push 2025-07-20 |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hosted on your own Linux x86_64 KVM serversFree, open source (Apache-2.0); self-hosted on your own Linux x86_64 KVM servers (Kubernetes or without it), so infrastructure cost is yours; no hosted service sold | N/a (self-hosted) | Pause keeps VM memory resident on the worker (not durable)Pause keeps VM memory resident on the worker (not durable); suspend snapshots memory, runtime state and root filesystem and releases the VM; wake restores either; sandboxes created through the SPACE API suspend after 6 h without access or command activity (idle limit, not a maximum lifetime) | VM | Jul 2026 | Sandbox platform behind Perplexity Computer, open-sourced; HTTP API with server-sent events for command outputSandbox platform behind Perplexity Computer, open-sourced; HTTP API with server-sent events for command output; templates, warm pools, base snapshots; E2B compatibility wrapper so the E2B Python SDK works against it; egress profiles; incubating snapshots, checkpoint and rollback; written in Rust; local Kind cluster install; 2,117 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0), self-hosted only; no hosted service or prices publishedPricing page | N/a (self-hosted) | Each sandbox has a TTL (default 30m, max 24h in the sample config) and expiresEach sandbox has a TTL (default 30m, max 24h in the sample config) and expires automatically; TTL can be extended through the API; per-exec timeout up to 10m; no pause or hibernate documented | VM | May 2026 | Self-hosted sandbox server (single binary) for private cloud, on-prem or air-gappedSelf-hosted sandbox server (single binary) for private cloud, on-prem or air-gapped; ~28ms microVM snapshot restore claimed (certified Linux/KVM hosts); Python and TypeScript SDKs; REST API; live preview URLs (Docker provider, port 3000 only); React web dashboard and stacyvm tui; pool mode packs several users per VM; signed Execution Receipts and worker identity; StacyVM is layer L0 (live) of a planned seven-layer StacyOS whose other layers are 'designed'; GitHub StacyOS/stacyvm 9 stars, last push 2026-06-05 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-hosted with Docker ComposePricing page | N/a (self-hosted) | One sandbox per task; the result comes back as a git branchOne sandbox per task; the result comes back as a git branch; lifecycle limits not documented | Container | Nov 2026 | Self-hosted server and worker that accept agent tasks over REST and MCP (submit_task)Self-hosted server and worker that accept agent tasks over REST and MCP (submit_task), plan first and run after you reply 'approve', run Claude Code, Codex, Gemini or Cursor CLIs pinned in the worker image, and hand back a git branch; workspace tokens with admin and member roles; provider keys encrypted at rest; collects 'learning rows' from members; the superlite.cloud domain does not resolve (NXDOMAIN); repo created 2026-11-14, 0 stars; same author as the superlite Claude Code workflow (iha2/superlite, 296 stars) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2024 | Sandboxed code execution for AI agents, locally or on the cloud. Massively parallel, easy to extendSandboxed code execution for AI agents, locally or on the cloud. Massively parallel, easy to extend. Powering SWE-agent and more. (644 GitHub stars; MIT; owner SWE-agent) |
| Runtime / SDK | Runtime / SDK | Self-hosted free (MIT / Apache-2.0); Premium (Enterprise Edition) $399/org/monthSelf-hosted free (MIT / Apache-2.0); Premium (Enterprise Edition) $399/org/month; Temps Cloud, a hosted control plane for your own servers, $29/month in invite-only beta; Temps sells no compute | N/a (self-hosted) | Idle timeout per sandbox (default 1 hour, max 24 hours): the container stops and keepsIdle timeout per sandbox (default 1 hour, max 24 hours): the container stops and keeps its files; the next exec or file call wakes a workspace; nothing deletes a sandbox automatically | Container | Jan 2026 | Self-hosted deploy platform (one Rust binary) that bundles git-push deploys, analytics, session replay, error trackingSelf-hosted deploy platform (one Rust binary) that bundles git-push deploys, analytics, session replay, error tracking, uptime, email and AI agent sandboxes; sandboxes are one-off runs or persistent workspaces, driven by CLI or REST API; Node SDK shaped like @vercel/sandbox in the repo (not on npm); per-sandbox CPU and memory limits; encrypted agent secrets; MCP server registration; AI features labelled Experimental; GitHub gotempsh/temps 916 stars |
| Runtime / SDK | Runtime / SDK | Self-hosted open source; no hosted service or prices | — | Ephemeral sandbox per session, reaped on stopEphemeral sandbox per session, reaped on stop; running sandboxes are reattached after an orchestrator restart | Container | Aug 2026 | Self-hosted single-tenant runtime that runs Claude Code sessions through the Claude Agent SDKSelf-hosted single-tenant runtime that runs Claude Code sessions through the Claude Agent SDK; Warden sidecar injects the real credential at egress while the sandbox holds a decoy; live per-session egress allow/deny; rewind workspace or conversation to a past turn; HMAC-chained audit receipts; web console and Python SDK (terrarium-python); Apache-2.0; 1 GitHub star, all commits on 2026-08-02 |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2026 | Micro-VM sandbox for AI agents: isolated, gated and auditable, to prove what your agents did (42 GitHub stars; MITMicro-VM sandbox for AI agents: isolated, gated and auditable, to prove what your agents did (42 GitHub stars; MIT; owner Cygnux-Labs) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-host on your own KubernetesPricing page | N/a (self-hosted) | auto_stop_interval stops the sandbox after N seconds of inactivity (0 = never)auto_stop_interval stops the sandbox after N seconds of inactivity (0 = never); stop releases compute and keeps state; auto_delete_interval deletes after stop (-1 = off, the default); optional persistent volume; cold snapshots for persistent sandboxes | Container | Mar 2026 | All-in-one AIO sandbox: shell, files, browser with human handoff, MCP, HTTP/WebSocket proxy; CLI, REST API, Python SDKAll-in-one AIO sandbox: shell, files, browser with human handoff, MCP, HTTP/WebSocket proxy; CLI, REST API, Python SDK; usage metering and plan enforcement built in; hosted site treadstone-ai.dev returned Cloudflare error 1016 on 2026-11-14; 34 stars |
| Runtime / SDK | Runtime / SDK | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2025 | The open-source AISI toolkit for sandboxing agentic evaluations (35 GitHub stars; no license fileThe open-source AISI toolkit for sandboxing agentic evaluations (35 GitHub stars; no license file; owner UKGovernmentBEIS) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hosted, and you pay for the TDX CVM provider, DNSFree, open source (Apache-2.0); self-hosted, and you pay for the TDX CVM provider, DNS, OIDC, container registry and Postgres | N/a (self-hosted) | The CLI launches, stops, starts, updates and terminates Dev CVMsThe CLI launches, stops, starts, updates and terminates Dev CVMs; updates keep provider-managed named volumes; persistent SSH sessions (dtach); no automatic suspend documented | VM | Aug 2026 | For coding agents (Claude Code, Codex) and editors over locally verified aTLS tunnelsFor coding agents (Claude Code, Codex) and editors over locally verified aTLS tunnels; Console control plane with OIDC, RBAC, policy, audit and traffic logs; credentials injected at the proxy instead of stored in the sandbox; SLSA-verified releases planned; pre-1.0; 12 stars; concrete-security.com did not load |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); runs on your own KVM host. The company's SSH relay is freePricing page | N/a (self-hosted) | Self-hosted: forks share the template's memory copy-on-writeSelf-hosted: forks share the template's memory copy-on-write, KSM merges identical pages and virtio-balloon reclaims idle memory; full and incremental snapshots; live migration | VM | Mar 2026 | Boot a template VM once, then fork copies ('under 20ms' on the clone pageBoot a template VM once, then fork copies ('under 20ms' on the clone page; the homepage and README say about 160 ms with networking and exec); '100 forks use memory like 10'; virtio-block (raw, qcow2), virtio-net, virtio-fs, VFIO passthrough including GPUs; guest agent for exec over vsock; VMs up to 64 GB RAM; repo description: 'Minimal VMM for multi-tenant dev shells and serverless'; the site doesn't offer a hosted sandbox service; other MIT tools: latch multiplexer, mosh-go, mobile SSH app; 436 stars |
| Runtime / SDK | Runtime / SDK | Free to self-host (site says MIT; the GitHub repo has no license file)Free to self-host (site says MIT; the GitHub repo has no license file); managed cloud API on a waitlist ("private beta" per the blog) with no published prices | N/a (VM destroyed after each execution) | A fresh microVM per execution from a pre-warmed pool, destroyed when the run finishesA fresh microVM per execution from a pre-warmed pool, destroyed when the run finishes; nothing persists between runs; memory and timeout caps set per request in capability tokens (examples: 256 MB, 30 s) | VM | May 2026 | Policy-driven execution guard for AI-generated code: HMAC-SHA256 capability tokens scope filesystem pathsPolicy-driven execution guard for AI-generated code: HMAC-SHA256 capability tokens scope filesystem paths, network on/off, syscalls, memory and timeout per request; JSONL audit trail; structured JSON results; daemon over a Unix socket; adapters for LangChain, LlamaIndex, CrewAI and OpenAI per the launch blog; claims ~150 ms p50 with a warm pool and ~900 ms cold start; the PyPI package vela-agent and crate vela-engine the site says to install return 404; GitHub repo has 4 stars and no pushes since 2026-05-31 |
| Runtime / SDK | Runtime / SDK | Self-hosted open source; no hosted service or prices | — | MemBox is ephemeral, in memory, destroyed when the sandbox endsMemBox is ephemeral, in memory, destroyed when the sandbox ends; GitBox persists files to S3, MinIO or local storage | WebAssembly | Feb 2025 | Python and a virtual shell (tree-sitter-bash parser, 30+ builtins) on an in-memory filesystem; isomorphic-gitPython and a virtual shell (tree-sitter-bash parser, 30+ builtins) on an in-memory filesystem; isomorphic-git; AI-friendly edit builtin; LangChain and Deep Agents integrations; orchestrator plus workers routed through Redis; Docker Compose stack; Apache-2.0; 0 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2026 | Lightweight, secure Linux sandboxes for untrusted processes. Runs in the browser and on the server. (78 GitHub starsLightweight, secure Linux sandboxes for untrusted processes. Runs in the browser and on the server. (78 GitHub stars; Apache-2.0; owner capsulerun) |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT); self-hosted ("not a hosted SaaS")Free, open source (MIT); self-hosted ("not a hosted SaaS"); you pay for your own compute and model provider | N/a (self-hosted) | One workload per run from a fresh worktree or clone on its own branchOne workload per run from a fresh worktree or clone on its own branch; watchdogs reconcile lost processes and pods; finalization salvages work before teardown; each run ends in a pushed branch (optional PR) | Container | May 2026 | Runs coding-agent harnesses (Pi and Claude Code adapters) as workloads with live events, steering, cancellationRuns coding-agent harnesses (Pi and Claude Code adapters) as workloads with live events, steering, cancellation, spend and concurrency caps, recovery and Git delivery (GitHub App; GitLab and Azure DevOps behind a flag); HTTP API, CLI and UI; one operator or trusted team per deployment, single bearer credential, no RBAC; public demo instance app.warren.run (TLS error on 2026-11-14); warren.run redirects to the GitHub repo; 518 stars; Product Hunt listing; latest release v0.19.3 (2026-10-07) |
| Runtime / SDK | Runtime / SDK | $0.24 | No active-CPU charge while stopped; whether provisioned memory is billed whileNo active-CPU charge while stopped; whether provisioned memory is billed while paused is not stated; volumes $0.05/GB-month | Edge instances start on a request and stop when idle (pricing page: pausing and resumingEdge instances start on a request and stop when idle (pricing page: pausing and resuming in a few milliseconds); Wasmer SDK sandboxes are created and destroyed by the app that embeds them | WebAssembly | Sep 2026 | Homepage: 'Universal compute. Built for AI'; Wasmer SDK (2026-09-01) runs local WebAssembly sandboxes for AI agentsHomepage: 'Universal compute. Built for AI'; Wasmer SDK (2026-09-01) runs local WebAssembly sandboxes for AI agents inside Python, JS/browser or Rust apps, claims 0.1 ms sandbox creation; registry packages for Python, Node (Edge.js), PHP, Bash, clang, FFmpeg; not a full Linux VM and no Docker; Edge: SSH/SFTP, cron and TCP/UDP apps on Pro, 12 global regions, custom domains, databases; runtime MIT, 21,244 GitHub stars |
| Runtime / SDK | Runtime / SDK | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2025 | Wassette: A security-oriented runtime that runs WebAssembly Components via MCP (977 GitHub stars; MIT; owner microsoft) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0); self-hosted on one Linux hostPricing page | N/a (self-hosted) | Persistent box per agent; sessions outlive the connection (detached PTY with replay)Persistent box per agent; sessions outlive the connection (detached PTY with replay); boxes stay up across daemon restarts; states REQUESTED, PROVISIONING, BOOTSTRAPPING, READY, DRAINING, DELETING, DELETED, plus quarantine; no automatic suspend documented | Process | Oct 2026 | Driven from the shell: workbox CLI, in-box wb toolkit, REST and WebSocket API, SDK, MCP serverDriven from the shell: workbox CLI, in-box wb toolkit, REST and WebSocket API, SDK, MCP server; Claude Code and Codex in the agents rootfs; network modes none, egress and allowlist; default hard limits 2 GiB memory, no swap, 2 cores, 4,096 pids; median 842 ms to ready and 11.9 MiB per idle sandbox (own measurements); scoped API keys, signed webhooks, event log, hash-chained audit log, brokered secrets; plain JavaScript on Node.js 22 with SQLite; 17 stars; roadmap: a Hyperfide provider |
| Runtime / SDK | Runtime / SDK | Free, open source (AGPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jun 2026 | Fast, self-hostable sandboxes for agents (102 GitHub stars; AGPL-3.0; owner mv37-org) |
| Runtime / SDK | Runtime / SDK | Free, open source (Apache-2.0), self-host on any Linux box with KVMFree, open source (Apache-2.0), self-host on any Linux box with KVM; managed API in early access (sign-up form), no prices published | — | Each exec forks a fresh VM from a template snapshotEach exec forks a fresh VM from a template snapshot; execution timeout default 30 s (timeout_seconds); no persistent sandboxes described; template update needs a full re-snapshot (~15 s) | VM | Mar 2026 | Spawn p50 0.79 ms, p99 1.74 ms, ~265 KB memory per sandbox, Python fork+exec ~8 msSpawn p50 0.79 ms, p99 1.74 ms, ~265 KB memory per sandbox, Python fork+exec ~8 ms, 1000 concurrent forks in 815 ms (their benchmarks); Python and TypeScript SDKs; REST API with public demo key; Prometheus metrics; limits: single vCPU per fork, no networking inside forks (serial I/O only), userspace PRNGs need reseeding per fork; written in Rust; 2,478 GitHub stars; zeroboot.dev redirects to the GitHub repo |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | No lifecycle of its own: the policy applies to the wrapped agent for as long as it runs | Process | Feb 2026 | macOS only; a single Bash script or Homebrew, no other dependenciesmacOS only; a single Bash script or Homebrew, no other dependencies; read/write to the selected workdir (git root by default), read-only toolchains, most of $HOME denied; profiles tested with Claude Code, Codex, OpenCode, Amp, Copilot CLI, Gemini CLI, Aider, Goose, Auggie, Pi, Cursor Agent, Cline, Kilo Code and Droid; --add-dirs/--add-dirs-ro and appended profiles; Git worktree detection; web policy builder; "a hardening layer, not a perfect security boundary"; GitHub eugene1g/agent-safehouse, 2,150 stars |
| Local | Local | Free, open source (the README badge says GPL-2.0Free, open source (the README badge says GPL-2.0; the BPF program declares Dual BSD/GPL and the repo has no LICENSE file) | N/a (runs on your own machine) | No lifecycle of its own: the ruleset binds the agent and every child process for as longNo lifecycle of its own: the ruleset binds the agent and every child process for as long as they run | Process | Jun 2026 | Confines the oh-my-pi coding agent (omp) to one project directoryConfines the oh-my-pi coding agent (omp) to one project directory; live terminal dashboard of blocked escape attempts and file opens; --audit runs the agent unconfined to show what it reaches for; --headless streams reports; strips secret-looking env vars (KEY, TOKEN, SECRET, AWS_); needs Linux 5.13+ with Landlock and BTF, plus the yeet daemon to load the BPF programs; sibling project agent-lock uses the BPF LSM instead; network exfiltration is not prevented; verified only on Linux 6.12 arm64 (Debian 13); by yeet (yeet.cx); 7 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Lightweight and declarative sandboxing for AI agents on Linux and macOS using Nix. (210 GitHub stars; MITLightweight and declarative sandboxing for AI agents on Linux and macOS using Nix. (210 GitHub stars; MIT; owner archie-judd) |
| Local | Local | Free, open source (MIT); runs on your own Linux machinePricing page | N/a (runs on your own machine) | No sandbox lifecycle: the agent starts and stops a hidden workspaceNo sandbox lifecycle: the agent starts and stops a hidden workspace; managed terminal sessions end when their workspace stops; a viewer can pause, set read-only or stop the agent | — | May 2026 | Rust MCP server (stdio, 82 tools) that gives a local agent its own hidden desktop and workspace-owned Chrome overRust MCP server (stdio, 82 tools) that gives a local agent its own hidden desktop and workspace-owned Chrome over loopback CDP, so it doesn't take over your mouse, focus or browser; six backends (X11, i3, KWin, GNOME, Hyprland, COSMIC); floating GPUI viewer with live pause, read-only and stop; optional daemon-enforced app allowlist; tools marked read-only, write or destructive; saved workspace profiles; bundled skill for Claude Code and Codex; npm, cargo or install.sh; Linux only, pre-1.0; GitHub 220 stars; sibling project computer-use-linux drives the real desktop |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2026 | Run multiple agents in parallel sandboxed VMs, with a single command, on your PC or in the cloud (581 GitHub starsRun multiple agents in parallel sandboxed VMs, with a single command, on your PC or in the cloud (581 GitHub stars; MIT; owner madarco) |
| Local | Local | Free, open source (Apache-2.0): hub, nodes, desktop app and CLIFree, open source (Apache-2.0): hub, nodes, desktop app and CLI; no account or subscription; you pay your model providers | N/a (runs locally) | Self-hosted; STOP ALL cancels every agent's running work and holds new work until resumeSelf-hosted; STOP ALL cancels every agent's running work and holds new work until resume, keeping queued work, VMs and workspaces; kill-and-resume tested; no idle timeout stated | VM | Oct 2026 | Self-hosted multi-agent orchestrator (orchestrator, planner, board) that runs each agent in its own microVMSelf-hosted multi-agent orchestrator (orchestrator, planner, board) that runs each agent in its own microVM; every tool call checked against owner-written rules ("laws"), with risky calls held for approval; budgets in dollars or tokens per project; model gateway on the hub adds provider keys from a vault so no key enters a VM; 12 coding engines inside VMs (Claude Code, Codex, OpenCode, Gemini CLI, Qwen Code, Cline, Goose and more); signed record of changes; runs on Windows, macOS, Linux and Raspberry Pi, with extra servers joined as nodes; Rust; 6 GitHub stars; built by JumpHash (NVIDIA Inception member) |
| Local | Local | Free, open source (GPL-3.0); runs on your own Linux KVM or Apple-silicon MacPricing page | N/a (runs locally) | One VM per project and agent, built once and reused; `stop` and `kill` end itOne VM per project and agent, built once and reused; `stop` and `kill` end it; the agent home volume (/home/dev) survives VM recreation until an explicit volume reset or prune | VM | Aug 2026 | Local launcher for opencode, Claude Code, Codex, pi, Hermes Agent and Oh My Pi inside a microVMLocal launcher for opencode, Claude Code, Codex, pi, Hermes Agent and Oh My Pi inside a microVM; host-side secret injection (the VM sees a placeholder); project directory mounted at /workspace, other host files hidden; network egress profiles (public, private, host, none) with allow/deny rules; optional VM-internal worktree sessions; links opened in the host browser by allowlist; Go; 88 GitHub stars; releases from v0.1.0 (2026-08-25) to v0.3.0; by inoio gmbh; match to the lead post is unconfirmed: 1996fanrui/agents-sandbox (agents-sandbox.com, Docker-based, 24 stars) has the same name |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Dec 2025 | Execution-Layer Security (ELS) for AI agents — policy-enforced shell with audit. (410 GitHub stars; Apache-2.0Execution-Layer Security (ELS) for AI agents — policy-enforced shell with audit. (410 GitHub stars; Apache-2.0; owner canyonroad) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2026 | Personal cloud OS for a developer and their AI agents on a single serverPersonal cloud OS for a developer and their AI agents on a single server. One-command install on Ubuntu, then everything in the browser: a windowed desktop, isolated workspaces with VS Code, Claude Code and Codex, a store of 944 self-hosted apps, backups and updates. Access only via Tailscale, nothing exposed to the internet. Rust core, Svelte UI. (942 GitHub stars; Apache-2.0; owner agentverse-os) |
| Local | Local | Free, open source (GPL-3.0-only)Pricing page | N/a (runs on your own machine) | Lasts as long as the wrapped agent; the private home is a fresh tmpfs per launch andLasts as long as the wrapped agent; the private home is a fresh tmpfs per launch and Linux /tmp is discarded on exit; only explicitly mounted state (agent state, --rw-map) persists; the approved per-project access policy is saved under ~/.config/ai-jail/projects | Process | Mar 2026 | Rust single binary with no daemon or root; Homebrew, AUR, .deb, crates.io and NixRust single binary with no daemon or root; Homebrew, AUR, .deb, crates.io and Nix; knows where Claude Code, Codex, Gemini CLI, OpenCode, Antigravity, Kimi, Kiro, Devin, Crush, Grok, pi, aider and others keep their login; first-run terminal UI with Fully closed, For developers and Mostly open presets; general internet closed except the agent's API host and package registries (Linux); phantom credentials (--secret), host port forwarding and publishing; isolated browser profiles; audit log; project secret masking; the README says it prevents accidents and is not a boundary against hostile code; by Fabio Akita; 1,507 GitHub stars |
| Local | Local | Free, open source CLI run on your own machine (the repo has no license file and GitHubFree, open source CLI run on your own machine (the repo has no license file and GitHub shows no license) | N/a (runs locally) | Disposable by default: the container is removed when you exitDisposable by default: the container is removed when you exit, and only the host project directory mounted at /workspace keeps changes; optional persistent devboxes keep packages and configuration across sessions until you delete them | Container | Sep 2026 | One dependency-free Bash script; running airlock in a project starts a preconfigured rootless Podman container with theOne dependency-free Bash script; running airlock in a project starts a preconfigured rootless Podman container with the project at /workspace; 26 preconfigured runtimes across languages, Linux distributions, security scanners and AI coding agents (OpenCode); .airlock project file for runtime, ports, env vars and setup hooks; installer also installs Podman; Linux, macOS and Windows via WSL2; 109 GitHub stars, 54 forks |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2025 | 🍺 Alcoholless: lightweight security sandbox for Homebrew, AI agents, etc. on macOS (208 GitHub stars; Apache-2.0🍺 Alcoholless: lightweight security sandbox for Homebrew, AI agents, etc. on macOS (208 GitHub stars; Apache-2.0; owner AkihiroSuda) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Dec 2025 | Amazing Sandbox - run third-party tools and AI agents securely on your machine (205 GitHub stars; MIT; owner ashishb) |
| Local | Local | Free, Apache-2.0 (npm @anthropic-ai/sandbox-runtime)Pricing page | — | N/a (wraps a local process; no lifecycle of its own) | Process | Oct 2025 | CLI srt and library; filesystem read/write rules, domain allowlist through HTTP/SOCKS proxiesCLI srt and library; filesystem read/write rules, domain allowlist through HTTP/SOCKS proxies, Unix socket restrictions; macOS violation monitoring; built for Claude Code, also for MCP servers and arbitrary commands; 7,125 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2025 | A tool for creating and running Linux containers using lightweight virtual machines on a MacA tool for creating and running Linux containers using lightweight virtual machines on a Mac. It is written in Swift, and optimized for Apple silicon. (51477 GitHub stars; Apache-2.0; owner apple) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Easily create full virtual machines that are sandboxed for development or computer use models. (1017 GitHub stars; MITEasily create full virtual machines that are sandboxed for development or computer use models. (1017 GitHub stars; MIT; owner jamesstringer90) |
| Local | Local | Free; build from source (no pre-built binaries yet); no hosted servicePricing page | N/a (runs on your own machine) | Each agent runs in an ephemeral containerEach agent runs in an ephemeral container; the agent is paused while a transaction waits for the user's approval | Container | Nov 2025 | Electron desktop app for AI agents that act on blockchains: agents never hold private keys and request signaturesElectron desktop app for AI agents that act on blockchains: agents never hold private keys and request signatures through @assegailabs/sdk (MIT, 1 star), with an approval popup per transaction and spending allowances; wallet through WalletConnect or a local Anvil test account; EVM chains (Ethereum, Polygon, Base, Optimism, Arbitrum) and testnets; a local LLM proxy adds the OpenAI or Anthropic key so the agent never sees it; the docs say to clone AssegaiLabs/Assegai-Agent-Sandbox, which is not public (GitHub returns 404) |
| Local | Local | Free, open source (MIT); runs on your own Mac or Linux machineFree, open source (MIT); runs on your own Mac or Linux machine, and exe.dev or boat.dev bill their own VMs | N/a (self-hosted; cloud backends bill their own VMs) | VMs stop and start keeping their disk; snapshots and forks copy disk state onlyVMs stop and start keeping their disk; snapshots and forks copy disk state only, never memory (about 1 s live on libvirt; Lima stops and restarts the VM); a shared pool hands out VMs under leases that expire and are renewed | VM | Oct 2026 | Lets coding agents (Claude Code, Codex, Cursor, opencode, pi) create, forkLets coding agents (Claude Code, Codex, Cursor, opencode, pi) create, fork, snapshot and remove VMs over MCP or a pi extension; every VM is a full XFCE desktop with Chrome for computer use and Playwright (desktop VM in about 13 s on libvirt) or headless; credential proxy, so VMs hold only a revocable token; grants for LLM keys, GitHub, ssh-agent and env vars; one MCP server drives several hosts at once; ephemeral GitHub Actions runners; 115 GitHub stars; sponsored by Modem |
| Local | Local | Free, open source (Apache-2.0); runs on your own Mac (Apple silicon) or Linux host withFree, open source (Apache-2.0); runs on your own Mac (Apple silicon) or Linux host with KVM | N/a (runs locally) | Sandbox keeps running after the agent exits so the next run starts at onceSandbox keeps running after the agent exits so the next run starts at once; `brig stop` stops it and keeps the guest home; `brig rm` deletes it; a change of project recreates the sandbox | VM | Aug 2026 | Runs Claude Code, Codex, Gemini, Grok or opencode in a local microVM that sees only the named project and its own homeRuns Claude Code, Codex, Gemini, Grok or opencode in a local microVM that sees only the named project and its own home; egress policies enforced outside the guest (gateway on macOS, nftables on Linux); credential broker gives the agent a session token and adds the real key per allowed request; signed guest images; bring your own OCI image as a profile; Go CLI (brig) and session daemon (brigd); pre-1.0; 266 GitHub stars; from the team behind urunc (CNCF Sandbox project) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Proper sandboxing for agentic coding and web browsing (307 GitHub stars; MIT; owner rderaison) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Run AI Agent CLIs in your browser (163 GitHub stars; Apache-2.0; owner leaningtech) |
| Local | Local | Free, open source (MIT); runs on your own machinePricing page | N/a (runs locally) | Records every step with a microVM snapshot (VM pausedRecords every step with a microVM snapshot (VM paused, CPU and memory state written to disk, resumed) so a sandbox can be rewound to any step; sandbox lives for the Python context or CLI session; no idle timeout documented | VM | Mar 2026 | 'Time-travel debugging for AI agent sandboxes': record, restore to step N, diff and rank agent sessions'Time-travel debugging for AI agent sandboxes': record, restore to step N, diff and rank agent sessions; 'bunkervm watch' installs a Claude Code PostToolUse hook and 'bunkervm review' flags deleted or skipped tests; MCP server (Claude Desktop, VS Code Copilot), CLI and Python API; LangChain and OpenAI Agents SDK adapters; Python 3.10+, pip install bunkervm (v0.14.0); GitHub 1 star |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Sandbox any macOS app — only your project directory stays accessible (109 GitHub stars; MIT; owner holtwick) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jun 2025 | A thin protective layer for Claude Code (429 GitHub stars; MIT; owner nikvdp) |
| Local | Local | Free, open source (AGPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2025 | Run coding agents inside Tart VMs on macOS [from repo/listing] (49 GitHub stars; AGPL-3.0; owner cirruslabs) |
| Local | Local | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Run AI coding agents in hardened container sandboxes. (131 GitHub stars; NOASSERTION; owner 89luca89) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2025 | ⚠️ development has moved to VibePod (174 GitHub stars; MIT; owner nezhar) |
| Local | Local | Free; public GitHub repo without a license file; runs on your own MacPricing page | N/a (runs on your own machine) | VMs run until you stop them (clawctl stop/start)VMs run until you stop them (clawctl stop/start); config and data live in a git-tracked project directory on the Mac, mounted into the VM over virtiofs, and survive VM rebuilds; snapshots by limactl clone of a stopped VM (APFS copy-on-write) | VM | Mar 2026 | macOS on Apple Silicon (M1-M4) only; interactive wizard or headless JSON configmacOS on Apple Silicon (M1-M4) only; interactive wizard or headless JSON config; Node.js, Tailscale and the 1Password CLI preinstalled in the VM; secrets as op:// and env:// references; several gateways side by side, each in its own VM; CPU and memory set per instance; 15+ AI providers; GitHub TimBeyer/clawctl 16 stars; latest release v0.23.0 (2026-05-13) |
| Local | Local | Free, open source (Apache-2.0); runs on your own machinePricing page | N/a (runs locally) | Idle VMs release memory down to about 1 GiB and stop after 30 idle minutesIdle VMs release memory down to about 1 GiB and stop after 30 idle minutes; clawk down stops (VM disk rebuilt fresh at next boot, repo and agent conversations persist on the host); clawk pause/resume in memory; clawk snapshot saves RAM and disk so resume restores running processes; destroy removes the VM | VM | Jul 2026 | One command boots a VM for the current repo and attaches Claude Code (or Codex, piOne command boots a VM for the current repo and attaches Claude Code (or Codex, pi, shell) with permission bypass flags; any OCI image as rootfs, no Docker needed; outbound traffic denied by default with a DNS-aware allow-list enforced below the guest, denials logged; ssh-agent forwarded so keys stay on the host; ticket mode with a git worktree per repo and cross-linked PRs; opt-in nested KVM for Docker or Kind inside; copy-on-write disks; 1,044 GitHub stars. The input URL github.com/lordrip/clawk is a fork (0 stars) of this repo |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | ClawLess — A serverless browser-based runtime for Claw AI Agents powered by WebContainers (545 GitHub stars; MITClawLess — A serverless browser-based runtime for Claw AI Agents powered by WebContainers (545 GitHub stars; MIT; owner open-gitagent) |
| Local | Local | Free, open source (Apache-2.0); runs on your own MacFree, open source (Apache-2.0); runs on your own Mac. README notes Tart is FSL-1.1-ALv2: free for individuals, commercial use may require a paid license | N/a (runs locally) | Base VM plus named instances (APFS copy-on-write clones) persist between sessionsBase VM plus named instances (APFS copy-on-write clones) persist between sessions; herdr agent sessions keep running after detach; clod stop stops all VMs, clod destroy removes named VMs | VM | Sep 2025 | Runs Claude Code, Codex, Gemini and Cursor Agent in a macOS VM with permission prompts disabledRuns Claude Code, Codex, Gemini and Cursor Agent in a macOS VM with permission prompts disabled; Xcode and dev tools preinstalled; bundled herdr agent multiplexer; headless mode (--no-graphics) for CI; three-layer APFS CoW caching for fast rebuilds; optional per-project isolation; shared memory budget (5/8 of host RAM by default); port forwarding over SSH reverse tunnels; per-project Brewfiles; 197 GitHub stars |
| Local | Local | Free, open source (MIT); runs on your own machine; you pay your own AI providerPricing page | — | — | Container | Jun 2026 | Sandbox for Claude Code, Codex, Gemini and Grok CLIs: does not mount SSH keys, cloud loginsSandbox for Claude Code, Codex, Gemini and Grok CLIs: does not mount SSH keys, cloud logins, host Git credentials or other repos; .env contents hidden, coop check-secrets scanner, .coopignore; network limited to the AI provider and approved destinations; stand-in API key inside the box; about 3 s to start once the image is built; task queue orchestrator for overnight agent work; Compose services run beside the box; MIT, 54 GitHub stars; by Protectorate |
| Local | Local | Free, open source (BSD-3-Clause)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Persistent Linux workspaces for agentic development. (391 GitHub stars; BSD-3-Clause; owner aerovato) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Give the agent a machine. Just not yours. Each AI coding agent gets its own isolated machine with root, DockerGive the agent a machine. Just not yours. Each AI coding agent gets its own isolated machine with root, Docker, and systemd - active defense detects and stops threats automatically. (786 GitHub stars; MIT; owner coipond) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2025 | MCP server running code in sandboxed containers [from repo/listing] (206 GitHub stars; MIT; owner philschmid) |
| Local | Local | Free, open source (AGPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2025 | A desktop app for isolated, parallel agentic development (2077 GitHub stars; AGPL-3.0; owner coder) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jun 2025 | A local sandbox for your AI agents (903 GitHub stars; Apache-2.0; owner instavm) |
| Local | Local | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2025 | Automate anything with Codex CLI in a local Docker container with cron, file watchers, webhooks, search/indexingAutomate anything with Codex CLI in a local Docker container with cron, file watchers, webhooks, search/indexing, speech, and hundreds of tools on tap. (112 GitHub stars; NOASSERTION; owner DeepBlueDynamics) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Give the agent a machine. Just not yours. Each AI coding agent gets its own isolated machine with root, DockerGive the agent a machine. Just not yours. Each AI coding agent gets its own isolated machine with root, Docker, and systemd - active defense detects and stops threats automatically. (786 GitHub stars; MIT; owner coipond) |
| Local | Local | Free, open source (Apache-2.0); runs on your own machinePricing page | N/a (runs locally) | Environments (container, git branch, command history) persist until you merge or applyEnvironments (container, git branch, command history) persist until you merge or apply the work and run container-use delete | Container | Jun 2025 | MCP server (container-use stdio) for Claude Code, Cursor, Goose, VS Code and other MCP agentsMCP server (container-use stdio) for Claude Code, Cursor, Goose, VS Code and other MCP agents; parallel agents without conflicts; full command history and logs; container-use terminal to drop into an agent's container; review with git checkout, then merge or apply; secrets management; Homebrew and install script; 4,093 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2026 | Isolated VM environment for running Claude Code and Codex (816 GitHub stars; Apache-2.0; owner trailofbits) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Apr 2026 | Sandbox for AI coding agents. Runs Copilot CLI, Claude Code, OpenCode, Gemini CLI, Antigravity, PiSandbox for AI coding agents. Runs Copilot CLI, Claude Code, OpenCode, Gemini CLI, Antigravity, Pi, goose or a plain shell inside a kernel-level sandbox, with git and gh guards and sandbox policy committed to the repository. (149 GitHub stars; MIT; owner navikt) |
Cua (Cua Spaces, Cua SDK)Hosted Cua Cloud and Fleets shut down 2026-10-18 Local | Local | Cua does not host or rent computers. Cua Spaces: Free $0 (2 computers in use at onceCua does not host or rent computers. Cua Spaces: Free $0 (2 computers in use at once, 20 remote relay hours/mo); Pro $18/mo or $180/yr (5 computers, 200 remote hours); Teams $30/seat/mo or $25/seat/mo yearly; VMs in your own cloud are billed by your cloud provider. The former Cua Cloud / Fleet hourly rates were removed | No hosted compute; your own hardware or cloud bill | Suspend and resume work only on local sandboxesSuspend and resume work only on local sandboxes; ephemeral sandboxes are deleted when their block exits; sandboxes in your own cloud (AWS, Google Cloud, Modal) get a time limit and are cleaned up. Hosted Cua Cloud (with cloud suspend/resume) ended 2026-10-18 | VM | Apr 2025 | Computer-use focus: Cua Driver lets agents use native macOSComputer-use focus: Cua Driver lets agents use native macOS, Windows and Linux apps without taking over the cursor (MIT); Lume macOS VM tool (MIT); Cua Spaces app (FSL-1.1-MIT, becomes MIT after 2 years) streams desktops over QUIC with a hosted relay, shared Spaces with per-person cursors, Teleport moves a signed-in app into a Space; agents: Claude Code, Codex, Cursor, Gemini CLI, Cline, Kiro, OpenClaw via MCP; Cua-Bench benchmark; repo trycua/cua 32,353 GitHub stars; Windows app in preview |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2025 | Development environments for coding agents. Enable multiple agents to work safely and independently with your preferredDevelopment environments for coding agents. Enable multiple agents to work safely and independently with your preferred stack. (4093 GitHub stars; Apache-2.0; owner dagger) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2026 | Lightweight, agent-native sandboxing with diff, commit and abort (329 GitHub stars; Apache-2.0; owner delta-box) |
| Local | Local | Free, open source (MIT); runs on your own machinePricing page | N/a (runs locally) | Persistent containers survive between sessions and resume where you left offPersistent containers survive between sessions and resume where you left off; per-container session counting stops devc stop while sessions are attached; optional stopOnLastDetach; devc down stops and removes, devc clean removes stopped containers | Container | Mar 2026 | Devcontainer.json compatible, with AI-safety settings under customizations.devcDevcontainer.json compatible, with AI-safety settings under customizations.devc; three security profiles: strict (no network, 2 CPU / 4 GB), moderate (domain allowlist, 4 CPU / 8 GB, default), permissive (host network); built-in profiles for Claude Code, Codex, Gemini CLI and Opencode with config mounts and network allowlists; macOS and Linux binaries (amd64, arm64) and Homebrew tap; README install links point to github.com/grahambrooks/devc, which returns 404; 2 GitHub stars |
| Local | Local | Free, open source (MIT); runs on your own machinePricing page | N/a (runs locally) | Per-project sandbox state (home overlay, tool caches) persists between runsPer-project sandbox state (home overlay, tool caches) persists between runs; Docker containers are kept after exit for 1-2 s restarts; --rm makes a sandbox ephemeral; persistent named scratchpads | Container | Feb 2026 | Brings your shell, aliases, mise-managed tools and editor configs into the sandbox while SSH keys, cloud credentialsBrings your shell, aliases, mise-managed tools and editor configs into the sandbox while SSH keys, cloud credentials, .env files and sibling projects stay out; .git read-only by default; optional MITM proxy logs, filters and redacts secrets in HTTP/HTTPS traffic, fail-closed egress on bwrap and krun; credential injection; shell wrappers so claude, codex, copilot, pi and opencode run sandboxed by default; sub-second start on Linux; krun defaults to 4 GB memory and 2 CPUs; 15 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); runs on your own machine (price per hour for 2 vCPU / 4Free, open source (Apache-2.0); runs on your own machine (price per hour for 2 vCPU / 4 GiB: n/a (runs on your own machine)) | — | Dexbox up boots a VM, dexbox stop stops the tool server; no suspend policy | VM | Mar 2026 | Go CLI and HTTP tool server that any agent (Claude, GPTGo CLI and HTTP tool server that any agent (Claude, GPT, Gemini) calls with tool actions in the model's native format: screenshot, mouse, keyboard, PowerShell; built-in MCP server for Cursor and Claude Code; browser viewer; shared folder between host and VM; reference agents in TypeScript (Vercel AI SDK) and Python (LangChain); macOS, Linux or Windows hosts; made by the Nen team; the HN link github.com/getdexbox/dexbox now returns 404; 9 GitHub stars; last push 2026-04-21 |
| Local | Local | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Dec 2016 | Android in docker solution with noVNC supported, video recording, mcp server and AI-agent (16136 GitHub starsAndroid in docker solution with noVNC supported, video recording, mcp server and AI-agent (16136 GitHub stars; NOASSERTION; owner budtmo) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2024 | Windows inside a Docker container. (54195 GitHub stars; MIT; owner dockur) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jul 2025 | Linux sandboxing that doesn't get in your way (405 GitHub stars; Apache-2.0; owner wrr) |
| Local | Local | Free, open source (MIT)Pricing page | N/a (self-hosted) | A foreground session's container stops when its terminal closesA foreground session's container stops when its terminal closes; background sessions can be detached and reattached; logins, config, history and agent memory persist on the host across sessions | Container | Jul 2026 | Runs Claude Code, Codex, Copilot, OpenCode, Pi (more via extensions) with no confirmation promptsRuns Claude Code, Codex, Copilot, OpenCode, Pi (more via extensions) with no confirmation prompts; all egress through an allowlist gateway with audit log; secrets reach the agent only as placeholders the gateway fills in; Theia IDE attach; HomeShell web/desktop UI (preview); Linux, macOS, Windows via WSL2; .deb/.rpm and self-updating binaries; Eclipse Foundation project; MIT; 80 stars |
| Local | Local | Free, open source (GPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Sep 2025 | A fully-featured, GUI-powered local LLM Agent sandbox with complete MCP protocol supportA fully-featured, GUI-powered local LLM Agent sandbox with complete MCP protocol support. Features both CLI and full desktop environment, enabling AI agents to operate browsers, terminal, and other desktop applications just like humans. Based on E2B oss code. (211 GitHub stars; GPL-3.0; owner BIGPPWONG) |
| Local | Local | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2026 | A personal VM cloud in one Go binary: persistent Linux VMs on macOS, Linux and WindowsA personal VM cloud in one Go binary: persistent Linux VMs on macOS, Linux and Windows, AI coding agents working inside them, any port published to HTTPS through Cloudflare Tunnel. Driven from a Mac OS 9 desktop in the browser, over SSH, or by an agent. (188 GitHub stars; no license file; owner livid) |
| Local | Local | Free, open source (MIT); runs on your own MacPricing page | N/a (runs locally) | VMs stop and start on demand (GUI or headless CLI with timeouts)VMs stop and start on demand (GUI or headless CLI with timeouts); snapshots and clones only of stopped VMs (APFS copy-on-write); no machine-state save/restore with the Custom VirGL graphics path | VM | Dec 2022 | Repo description: isolated macOS or Linux VMs on Apple Silicon for AI agents to safely run code inRepo description: isolated macOS or Linux VMs on Apple Silicon for AI agents to safely run code in; README and site describe a general-purpose VM manager and list local automation for scripts and AI agents as a direction, with a local API/MCP surface in the backlog; ezvm CLI with versioned JSON output; authenticated Linux guest agent (IP reporting, file transfer, shutdown); Ubuntu, Debian, Fedora, Kali and macOS guests; needs macOS 27+ on Apple Silicon; signed and notarized, Homebrew cask; github.com/everettjf/vm4a redirects to everettjf/ezvm; 117 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | No lifecycle of its own: wraps one command or agent sessionNo lifecycle of its own: wraps one command or agent session, and the restrictions last as long as that process runs | Process | Dec 2025 | Go CLI and Go library; works with Claude Code, Codex, Amp, Gemini CLI, GitHub Copilot, OpenCodeGo CLI and Go library; works with Claude Code, Codex, Amp, Gemini CLI, GitHub Copilot, OpenCode, Factory Droid and other CLI agents; domain allowlists, filesystem read/write rules, command deny rules (e.g. git push, rm -rf /), SSH host and command filtering; built-in templates such as "code"; monitor mode logs violations (bpftrace optional on Linux); imports rules from Claude Code settings; agent hooks; macOS and Linux, WSL through Nix; inspired by Anthropic's sandbox-runtime; moved from Use-Tusk/fence; Show HN 2026-01-20; 1,022 GitHub stars |
| Local | Local | Free (MIT, local only)Pricing page | $0 | Disposable: the box is destroyed when the agent exits; a failed run freezes a snapshot | VM | Aug 2026 | Policy and UX layer over microsandbox; mounts only the repo at /workspace; network off or allowlisted from frag.tomlPolicy and UX layer over microsandbox; mounts only the repo at /workspace; network off or allowlisted from frag.toml; MCP the agent cannot widen; 1 GitHub star |
| Local | Local | Free, open source (Apache-2.0); runs on your own machinePricing page | N/a (runs on your own machine) | Each VM starts from a clean image and its root disk is thrown away on closeEach VM starts from a clean image and its root disk is thrown away on close; snapshotting a running session stops it and saves a qcow2 disk checkpoint that new VMs resume from (disk only; /tmp and other tmpfs paths are not kept); running sessions can be listed and re-attached | VM | Feb 2026 | TypeScript SDK and CLI (npx @earendil-works/gondolin, Node.js 22.19+); micro-VMs boot in under a second (docs)TypeScript SDK and CLI (npx @earendil-works/gondolin, Node.js 22.19+); micro-VMs boot in under a second (docs); programmable HTTP/TLS egress allowlists and request/response hooks in JavaScript; secrets injected by the host for allowed hosts, the guest sees only placeholders; programmable VFS mounts; ingress gateway; SSH into the guest and allowlisted SSH egress; custom Alpine image builds with optional OCI rootfs; remote machines through a static Rust daemon; macOS (Apple silicon, Intel) and Linux (x86_64, arm64); Pi extension and Browser Use example; by Earendil Works; 2,430 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Container-free, deny-by-default sandbox for AI coding agentsContainer-free, deny-by-default sandbox for AI coding agents. Kernel-enforced filesystem, network, and syscall isolation for Linux and macOS (315 GitHub stars; Apache-2.0; owner GreyhavenHQ) |
| Local | Local | Free, open source (Apache-2.0); runs locallyPricing page | N/a (local) | Boxes are git worktrees with states created, running/idle, proposed, applied or abortedBoxes are git worktrees with states created, running/idle, proposed, applied or aborted; kept until h5i box gc or rm; no auto-suspend stated | Process | Mar 2026 | Mainly an Axum-based Rust web framework with Lean 4 verificationMainly an Axum-based Rust web framework with Lean 4 verification; h5i box runs Claude Code or Codex in a local sandbox with runtime-scoped credentials, a host-side proxy that injects API keys, egress allowlists and export as patch + report + receipt; browser profile with headless Chrome; red-team and security-testing CLI; Apache-2.0; 924 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2026 | Hanzo VM — native microVMs for agents: Virtualization.framework on macOS, KVM on LinuxHanzo VM — native microVMs for agents: Virtualization.framework on macOS, KVM on Linux. Mounts, checkpoints, port forwards, egress policy, secret injection. (3 GitHub stars; Apache-2.0; owner hanzoai) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Containment for AI agents - user isolation, sandboxed execution, network controls, backup/rollback. TLA+ verifiedContainment for AI agents - user isolation, sandboxed execution, network controls, backup/rollback. TLA+ verified. (179 GitHub stars; MIT; owner dredozubov) |
| Local | Local | Free, open source (the README says MIT; the repo has no LICENSE file)Pricing page | N/a (runs on your own machine) | Each run gets a writable copy of the rootfsEach run gets a writable copy of the rootfs; containers are cleaned up automatically when the daemon restarts; the hopsd daemon starts on demand | VM | Jan 2026 | Swift CLI and daemon (gRPC over a Unix socket) plus an optional Rust GUI; macOS 26+ on Apple silicon onlySwift CLI and daemon (gRPC over a Unix socket) plus an optional Rust GUI; macOS 26+ on Apple silicon only; TOML profiles with allow/deny path lists; network modes disabled, loopback, outbound or full; CPU, memory and process-count limits; interactive TTY; a general sandbox for untrusted code: the README doesn't mention AI agents; 5 GitHub stars |
| Local | Local | Free, open source (MIT); runs on your own machinePricing page | N/a (runs locally) | Desktops persist: home directory and Chromium profile live on Docker volumesDesktops persist: home directory and Chromium profile live on Docker volumes; each tool call takes or renews a five-minute reservation that expires when idle; stop keeps the home; a checkpoint copies a stopped desktop's home (no running processes) for backup or cloning | Container | Sep 2026 | Agents reserve a desktop through MCP (cua-computer-server computer tools plus Playwright MCP)Agents reserve a desktop through MCP (cua-computer-server computer tools plus Playwright MCP); human watches or takes over in a noVNC viewer, then hands back; clones from checkpoints only with user approval; optional Valet credential and payment broker; max_running and memory budget; example desktop needs 2 GiB; macOS or Linux with a local Docker daemon; 7 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); runs locallyPricing page | N/a (runs locally) | Stateful IPython kernels keep definitions and variables across executions for the life ofStateful IPython kernels keep definitions and variables across executions for the life of the kernel; no idle timeout documented | Process | Dec 2024 | Runs code blocks that mix Python, shell commands and programmatic MCP tool calls ('code mode')Runs code blocks that mix Python, shell commands and programmatic MCP tool calls ('code mode'); generates typed Python APIs for MCP tools via mcpygen; per-call approval of tool calls and shell commands; Python SDK, MCP server, Claude Code plugin; the freeact agent is built on it; default sandbox blocks internet access; v0.9.2 on PyPI; GitHub gradion-ai/ipybox 76 stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Secure Nix sandbox for LLM agents (134 GitHub stars; MIT; owner andersonjoseph) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2026 | Secure agents in seconds by running them in a sandbox. (6 GitHub stars; MIT; owner kontext-security) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | Named VMs persist until krunai delete; krunai stop keeps the VM and its diskNamed VMs persist until krunai delete; krunai stop keeps the VM and its disk; clone, export to a tarball (VM must be stopped) and import | VM | Feb 2026 | Rust CLI; Homebrew tap for macOS and a Fedora COPR package, although the README lists Linux as "soonRust CLI; Homebrew tap for macOS and a Fedora COPR package, although the README lists Linux as "soon, waiting on a new libkrunfw release"; the project directory is shared into the VM at ~/work; you install the agent inside the VM yourself or with a setup script (examples in contrib/); SSH access; port forwarding (--port); defaults to 4 vCPUs and 8192 MB RAM (--cpus, --mem); by Sergio López; 50 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | N/a (runs on your own machine) | No lifecycle of its own: the restrictions apply to the wrapped process and its childrenNo lifecycle of its own: the restrictions apply to the wrapped process and its children for as long as they run | Process | Mar 2025 | Go CLI (landrun [options] <command>); needs Linux kernel 5.13+ (6.7+ for TCP rules, 7.2+ for UDP rules)Go CLI (landrun [options] <command>); needs Linux kernel 5.13+ (6.7+ for TCP rules, 7.2+ for UDP rules), --best-effort falls back on older kernels; denies all access unless paths or ports are allowed; passes no environment variables by default; audit logging of denials (ABI v7+); packaged for Arch (AUR), Slackware, Debian forky and Ubuntu 26.04; general-purpose: the README doesn't mention AI agents; Show HN 2025-03-22 (518 points); 2,341 GitHub stars |
| Local | Local | Free, open source: binary LGPL-2.1-or-laterFree, open source: binary LGPL-2.1-or-later; Node.js API, OpenCode plugin and Pi extension Apache-2.0 | N/a (runs on your own machine) | No lifecycle of its own: runs one command under a merged JSON policy for as long as itNo lifecycle of its own: runs one command under a merged JSON policy for as long as it runs | Process | Jun 2026 | Installed from npm (@landstrip/landstrip) and run as npx landstrip run -p policy.json -- <command>Installed from npm (@landstrip/landstrip) and run as npx landstrip run -p policy.json -- <command>; JSON policy with allowWrite/denyWrite/allowRead/denyRead, network on/off, local binding, HTTP and SOCKS proxy ports, Unix sockets and macOS- and Windows-specific keys; check and doctor commands; denial reporting over a trap descriptor; TypeScript API (@landstrip/landstrip-api); OpenCode plugin and Pi extension; repo description: "A sandbox for coding agents with parametrized state"; 91 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); runs on your own machinePricing page | N/a (runs on your own machine) | Agents keep working after the terminal closes and the next `lea` reattachesAgents keep working after the terminal closes and the next `lea` reattaches; `lea stop` stops them; no hosted lifecycle | Container | Oct 2026 | Runs many agents side by side in one container: Claude Code by default, CodexRuns many agents side by side in one container: Claude Code by default, Codex, or any OpenAI-compatible endpoint (Ollama, vLLM) with the key kept in the proxy; lea deck terminal UI and lea web; a Control Tower agent that starts and follows dev and reviewer agents; agents message each other and lease files; allowances for push, PR and merge; GitHub token usable but not readable; Lock Family brokers (locksql, lockmq, lockfile, locklog, lockssh, lockhttp) with approvals in `lox run`; cosign-signed releases; Homebrew install; 2 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2025 | Leash by StrongDM - take your AI agents for a walk (599 GitHub stars; Apache-2.0; owner strongdm) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | The container stops automatically when the last terminal exitsThe container stops automatically when the last terminal exits; the home directory lives on the host (~/Litterbox/homes), and changes outside it are committed when the container is recreated; litterbox delete removes a box | Container | Aug 2025 | Rust CLI for Linux (installer x86-64 only); each box is defined by a Dockerfile templateRust CLI for Linux (installer x86-64 only); each box is defined by a Dockerfile template; a custom SSH agent exposes chosen keys to one box and asks for confirmation on every signing request; device passthrough; optional PipeWire; Wayland and X11 (xwayland-satellite) apps run inside; meant to limit damage from supply-chain attacks on developers; the README and site don't mention AI agents; "still early" and not yet stable; 162 GitHub stars |
| Local | Local | Free, open source (GPL-3.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Nov 2025 | An MCP server that executes Python code in isolated rootless containers with optional MCP server proxyingAn MCP server that executes Python code in isolated rootless containers with optional MCP server proxying. Implementation of Anthropic's and Cloudflare's ideas for reducing MCP tool definitions context bloat. (344 GitHub stars; GPL-3.0; owner elusznik) |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2026 | Kernel-level eBPF sandbox for securing LLM agent tool calls made through the Model Context Protocol (MCP) (75 GitHubKernel-level eBPF sandbox for securing LLM agent tool calls made through the Model Context Protocol (MCP) (75 GitHub stars; MIT; owner meta) |
| Local | Local | Free, open source (MIT); an SDK built into your own appPricing page | N/a (runs locally) | One-shot create-and-run with a per-request timeout (timeoutMs)One-shot create-and-run with a per-request timeout (timeoutMs), or persistent containers that are provisioned, started, executed in, stopped and deprovisioned; no idle suspend | VM | Oct 2026 | Policy-driven containment for AI agents and untrusted code (model output, plugins, tools, the whole agent)Policy-driven containment for AI agents and untrusted code (model output, plugins, tools, the whole agent); JSON policies for filesystem (read-only, read-write, denied paths), network (proxy, egress controls) and UI (clipboard, display); Rust, .NET and Node SDKs (crates.io, NuGet, npm); audit mode to learn policies; Windows release builds send required telemetry to Microsoft; Windows 365 Cloud PC support GA; GitHub microsoft/mxc 3,142 stars |
| Local | Local | Free, open source (Apache-2.0), public betaFree, open source (Apache-2.0), public beta; runs on your own machine. Minimal for Teams (managed hermetic builds, org-wide policy, attestation) is announced as coming soon with no price | N/a (runs on your own machine) | Sessions are durable: they keep running after you detach, with no time limitSessions are durable: they keep running after you detach, with no time limit; stopping the provider (min stop) ends running shells but session records and workspaces persist and are re-hosted when it comes back; a session is removed only by min session destroy | VM | Aug 2025 | Local-first isolated dev and agent sessions declared in a minimal.toml that travels across branchesLocal-first isolated dev and agent sessions declared in a minimal.toml that travels across branches, forks and worktrees; Linux x86-64/ARM64 (kernel 5.10+) and macOS Apple Silicon (M1+), no Intel Macs; agent policies for file system and network (egress/ingress rules); loadouts bring your own editor and shell config; packages built from source per Nickel build specs, SLSA Build L3, CycloneDX SBOM and VEX; hosted MCP server and OpenAPI for the package catalog; Claude Code skills repo; remote execution coming soon; GitHub gominimal/minimal 240 stars |
| Local | Local | Free, open source (Apache-2.0); runs on your own MacPricing page | N/a (runs locally) | Ephemeral: instant APFS clones; mirage run clones, runs and destroysEphemeral: instant APFS clones; mirage run clones, runs and destroys; warm snapshot saves memory and disk together and restores to the logged-in desktop; host limit of 2 running macOS VMs (Apple's limit) | VM | Jun 2026 | Clone in about 10 ms regardless of disk size; MCP server with vm_clone, vm_exec, vm_run, vm_snapshotClone in about 10 ms regardless of disk size; MCP server with vm_clone, vm_exec, vm_run, vm_snapshot, vm_screenshot and more; host-guest control over vsock (no SSH); native SwiftUI app with live VM view; zero-touch golden image with auto-login; every command has --json; macOS guests only; Apple Silicon, macOS 14+; 2 GitHub stars; all commits on 2026-06-23 |
| Local | Local | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (Apache-2.0) | — | Mvm pause freezes a VM in memory at zero CPU, mvm resume continues itMvm pause freezes a VM in memory at zero CPU, mvm resume continues it; memory+disk snapshots and restore (restore broken on macOS 26.2 per README); no idle auto-suspend stated | VM | Apr 2026 | Local-first: native microVMs on Apple Silicon, same CLI and API on a self-hosted Linux server (install ~95 s)Local-first: native microVMs on Apple Silicon, same CLI and API on a self-hosted Linux server (install ~95 s); README: 0.697 s cold boot and 0.293 s restore on Apple VZ; exec over vsock; warm pool and network policy; Python (mvm-sandbox), TypeScript (@agentstep/mvm-sdk) and Go SDKs; Apache-2.0; 2 stars |
| Local | Local | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Self hosted cloud coding agent with k3s + kata containers + cloud hypervisor microVMs + tailscale + any harness + aSelf hosted cloud coding agent with k3s + kata containers + cloud hypervisor microVMs + tailscale + any harness + a nice iOS app (263 GitHub stars; no license file; owner angristan) |
| Local | Local | Free, open source (Apache-2.0) CLI (the site lists the price as $0)Free, open source (Apache-2.0) CLI (the site lists the price as $0); the nolabs platform for agent fleets is "in development" with no published price | N/a (runs on your own machine) | Restrictions last as long as the wrapped agent runsRestrictions last as long as the wrapped agent runs; each brokered tool invocation's sandbox is destroyed when the invocation exits; "Ghost Sessions" give a detachable session lifecycle; atomic filesystem snapshots allow undo and rollback | Process | Jan 2026 | macOS, Linux and Windows (WSL2); agent profiles in registry.nono.sh for Claude Code, Codex, Copilot, OpenCode, GoosemacOS, Linux and Windows (WSL2); agent profiles in registry.nono.sh for Claude Code, Codex, Copilot, OpenCode, Goose, OpenClaw, Hermes, Pi and others; proxy-based credential injection with phantom tokens; domain filtering and L7 rules on HTTP method and path; argv policy per tool with human approval for out-of-policy calls; Sigstore attestation of skills, hooks and instruction files; hash-chained audit log sealed with a Merkle root; Rust and Python bindings; runs on laptops, in CI, containers, Kubernetes pods or microVMs; "created by the team behind Sigstore"; pre-1.0 (v0.82.0, 2026-11-14); 4,986 GitHub stars |
| Local | Local | Free, open source (AGPL-3.0); runs on your own machinePricing page | N/a (runs on your own machine) | Spin one up per agent and tear it down when doneSpin one up per agent and tear it down when done; the per-agent overlay keeps that agent's state | VM | Oct 2026 | Virtual desktop OS built for agents, from the AnythingLLM team, which plans to bring it into AnythingLLMVirtual desktop OS built for agents, from the AnythingLLM team, which plans to bring it into AnythingLLM; XFCE themed like Windows 10 with a live UI the human watches; Chromium, terminal, editor and file manager; Pi agent harness with custom extensions and a memory manager; ask-user for human-in-the-loop; works without a vision model through accessibility APIs and claims about 60% fewer browser tokens than browser-use; any OpenAI-compatible model (LM Studio, Ollama, cloud); prebuilt QEMU and base image for macOS ARM64 and Windows x64; base image about 2.9 GB, per-agent overlay about 100 MB; passwordless sudo; 16 GitHub stars |
| Local | Local | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Oct 2025 | Sandboxed execution of commands in Docker containers with automated worktree and dev container management (175 GitHubSandboxed execution of commands in Docker containers with automated worktree and dev container management (175 GitHub stars; no license file; owner obra) |
| Local | Local | Free, open source (MIT); runs on your own machine (price per hour for 2 vCPU / 4 GiB: n/aFree, open source (MIT); runs on your own machine (price per hour for 2 vCPU / 4 GiB: n/a (runs on your own machine)) | — | Workspaces are long-lived containers on your own host; no suspend or timeout described | Container | Jan 2026 | Daemon (perry agent run) that creates workspaces and registers each on your Tailscale tailnet by nameDaemon (perry agent run) that creates workspaces and registers each on your Tailscale tailnet by name; CLI, web UI (port 7391) and SSH from any device on the tailnet; Claude Code, OpenCode and Codex CLI preinstalled, with credentials synced from the host; Ubuntu 24.04 workspaces with Node 22, Python 3, Go and Docker; mobile apps buildable from source; README warns the web UI and API have no authentication and must stay inside a private network; latest release v0.5.1 (2026-02-16); 37 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); no account; runs on your own computer and phonePricing page | N/a (runs on your own devices) | N/a (your own phone; Pinky's desktop app runs while openN/a (your own phone; Pinky's desktop app runs while open, and scheduled tasks can run at set times) | VM | Nov 2026 | Desktop app for macOS (Apple silicon and Intel), Windows and Linux that lets any MCP agent (Claude CodeDesktop app for macOS (Apple silicon and Intel), Windows and Linux that lets any MCP agent (Claude Code, Claude Desktop, Codex, Cursor) operate a real iPhone (iOS 17.4+) or Android phone (Android 10-17) over stdio or HTTP MCP; you watch the mirror, approve payments and other risky taps, take over at any time, and get a step log, screenshots and a screen recording per run; Pinky app on the phone can run tasks itself with a model you choose (Claude, GPT, Gemini, DeepSeek, Qwen, local Ollama); PIN unlock with Touch ID or Windows Hello; works over USB, Wi-Fi or Tailscale; 16-22 MB native app; 2,152 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); runs on your own MacPricing page | N/a (runs locally) | Containers start on demand with no poolingContainers start on demand with no pooling; stop, restart and remove by hand (sand does not yet watch activity to stop idle sandboxes) | Container | Sep 2025 | APFS copy-on-write clone of the project mounted at /app, so the real working directory stays on the hostAPFS copy-on-write clone of the project mounted at /app, so the real working directory stays on the host; --agent claude, codex, gemini or opencode; sand git status/diff to inspect drift, git pull from a sand remote to bring work back; network filtering; one-shot non-interactive runs; profiles; Apple Silicon, macOS 26+, Apple container CLI 1.7.1; 23 GitHub stars. The input URL github.com/erikswedberg/sand is a fork (0 stars) of this repo |
| Local | Local | Free, open source (MIT); runs on your own PC and needs Windows ProFree, open source (MIT); runs on your own PC and needs Windows Pro, Enterprise or Education with the Windows Sandbox feature (README: "free with Windows Pro/Enterprise") | N/a (runs on your own PC) | Fresh on every boot; the whole guest (files, registry keysFresh on every boot; the whole guest (files, registry keys, installed apps) is deleted when the sandbox stops; explicitly shared output folders, screenshots and reports stay on the host; no persistent environment; one sandbox at a time | VM | Nov 2026 | Local Windows only: sbx CLI plus optional MCP server sharing one engineLocal Windows only: sbx CLI plus optional MCP server sharing one engine; clean Windows 11 boots in about 15 s, GUI test samples take 40-50 s end to end; UI Automation for WinUI, WPF, WinForms, Win32 and MSIX apps; screenshots, recordings, HTML and MP4 reports; network off / firewall / on; dependency diagnosis and opt-in verified runtime provisioning; installer/uninstaller diffs; sbx detonate for untrusted files; portable Windows x64 ZIP with bundled Node.js, no ARM64 package; by Scott Hanselman; 432 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Mar 2026 | Orchestrate sandboxed coding agents in TypeScript with sandcastle.run() (9079 GitHub stars; MIT; owner mattpocock) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | A Docker & dev container setup for securely running AI agents in `--dangerous` modeA Docker & dev container setup for securely running AI agents in `--dangerous` mode. All container traffic is routed through a transparent mitmproxy, enforcing network access rules and injecting secrets. (204 GitHub stars; Apache-2.0; owner VirtusLab) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | The sandvault user account and its home directory persist between runsThe sandvault user account and its home directory persist between runs; commands run through sudo -u or SSH to that account; files in /Users/Shared/sv-$USER survive rebuilds and sv uninstall | VM | Sep 2025 | Homebrew (brew install sandvault); includes Claude Code, OpenAI Codex, OpenCode, Google Gemini, piHomebrew (brew install sandvault); includes Claude Code, OpenAI Codex, OpenCode, Google Gemini, pi, Muse Code and the herdr agent multiplexer; no access to your home directory, other users' directories or mounted drives; shared workspace /Users/Shared/sv-$USER; passwordless switching after setup; SSH mode; Chrome/Lightpanda browser automation and iOS Simulator access from inside the sandbox; xcodebuild and swift support; GUI apps can't run inside; by Patrick Wyatt; 454 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Aug 2025 | Build product with grounded, parallel coding agents. (246 GitHub stars; MIT; owner imbue-ai) |
| Local | Local | Free, open source (MIT)Pricing page | N/a (runs on your own machine) | Ephemeral: each session runs in a fresh auto-removed container ("Exit and it's gone") | Container | Nov 2025 | CLI via npm (@colony2/shai) or Homebrew; for Codex, Claude Code, Gemini CLI, Cline, OpenHands and other CLI agentsCLI via npm (@colony2/shai) or Homebrew; for Codex, Claude Code, Gemini CLI, Cline, OpenHands and other CLI agents; "cellular development": write access only to chosen subdirectories; resource sets (allowed hosts, mounts, ports, env vars, root commands) applied per path; selective elevation through host-side commands exposed by a host MCP server; shai-base and shai-mega images; Go embedding API; by colony2.com; 45 GitHub stars |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | A local-first microVM sandbox for running AI agents safely on macOS & Linux (894 GitHub stars; Apache-2.0A local-first microVM sandbox for running AI agents safely on macOS & Linux (894 GitHub stars; Apache-2.0; owner superhq-ai) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | One-shot: each invocation boots a fresh VM that ends with the command ("VMs die when theyOne-shot: each invocation boots a fresh VM that ends with the command ("VMs die when they exit"); no daemon and 0 MB idle memory; silo build bakes persistent rootfs layers that later runs reuse | VM | Apr 2026 | CLI via Homebrew (rchekalov/apps/silo); shims route python, node, cargo, go and deno (and claudeCLI via Homebrew (rchekalov/apps/silo); shims route python, node, cargo, go and deno (and claude, after silo install claude-code) through VMs; only the current directory is mounted read-write; network off by default with a per-project allowlist in silo.toml; tool registry includes python, node, rust, go, deno, psql and claude-code; ~600 ms warm start (583 ms p50 on M2 Pro), ~25 s on first run; 2 GB per VM by default (claude-code 4 GB); repo description: "Sandbox npm, pip, and AI coding agents in ephemeral macOS micro-VMs"; also served at silo.rchekalov.com; 8 GitHub stars |
| Local | Local | Free, Apache-2.0, runs on your own laptop or serversFree, Apache-2.0, runs on your own laptop or servers. Hosted smol cloud from the same team: $0.04/machine-hour base + $0.05/active vCPU-hour + $0.0162/GB-hour active memory + disk $0.0001/GB-hour; plans $0/$20/$150 per month + usage; $100 signup credit | — | Machines persist across restarts; pause/resumeMachines persist across restarts; pause/resume; checkpoints capture RAM, CPU state and disks; live copy-on-write branches; elastic memory via virtio balloon | VM | Dec 2025 | Boots <200 ms; any OCI image; .smolmachine portable artifacts (self-contained executable)Boots <200 ms; any OCI image; .smolmachine portable artifacts (self-contained executable); checkpoints move Mac arm64 to Linux arm64; SDKs for Node, Python, Rust, Go; network off by default, host allowlists, credential substitution; Kubernetes RuntimeClass; GPU via Vulkan/CUDA remoting; browser display; defaults 4 vCPU, up to 8 GiB; 7,100 GitHub stars; smol machines is backed by Y Combinator (smolmachines.com) |
| Local | Local | Free, open source (Apache-2.0)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | CLI tool for running coding agents inside hardware-isolated microVMs (80 GitHub stars; Apache-2.0; owner stacklok) |
| Local | Local | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Coding-agent VM orchestrator: runs coding agents in isolated VMs — microVMs on Linux (with ZFS-based audit-trailCoding-agent VM orchestrator: runs coding agents in isolated VMs — microVMs on Linux (with ZFS-based audit-trail snapshots) and Apple's container tool on macOS. (43 GitHub stars; no license file; owner prime-radiant-inc) |
| Local | Local | Free, open source (Apache-2.0); runs on your own Mac or Linux ARM64 machineFree, open source (Apache-2.0); runs on your own Mac or Linux ARM64 machine, or inside AgentCore Runtime or Lambda MicroVMs (which AWS bills separately) | N/a (runs locally) | Runs one agent workload per box on the host; no pause or snapshotRuns one agent workload per box on the host; no pause or snapshot; a box keeps its ID and can have its policy reloaded while running | Process | Oct 2026 | Harness-agnostic: run any agent binary, with access set in box.toml and rules in Dogwood (policy.dw)Harness-agnostic: run any agent binary, with access set in box.toml and rules in Dogwood (policy.dw); default-deny, temporal rules can depend on earlier actions and elapsed time; embedded Strands Shell and Monty for Python interpreters route file and command operations through the policy engine; egress gateway checks connections and HTTP method/path and injects API credentials or AWS SigV4 so the agent never holds secrets; MCP broker checks tool calls and arguments; decision records in OTLP JSON; usable as a Rust library; Linux x86-64 experimental; policy-authoring agent skill; 546 GitHub stars |
| Local | Local | Free, open source (MIT); runs on your own machinePricing page | N/a (runs locally) | Each run bounded by --max-duration=15m (with --memory=2g --cpus=1.0)Each run bounded by --max-duration=15m (with --memory=2g --cpus=1.0); workspace snapshot before each run, sd rewind restores tracked files and removes new ones | Container | Jul 2026 | Sd run wraps Claude Code, Cursor, OpenDevin or aider; git-plumbing workspace snapshots with sd rewind and sdSd run wraps Claude Code, Cursor, OpenDevin or aider; git-plumbing workspace snapshots with sd rewind and sd checkpoint; dummy-key substitution proxy for API credentials; sd swarm runs several agents on an internal Docker bridge network; sd record captures traffic as Keploy mocks; optional local control plane and dashboard; npm stuntdouble-sandbox-cli. The site claims eBPF interceptors that drop SQL DROP/DELETE and native macOS and Windows kernel interceptors, but the README says the eBPF program is never compiled, the native interceptors were never built, and policies are advisory; 1 GitHub star |
| Local | Local | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | A lightweight x86-64 / arm64 Virtual Machine Monitor (VMM) for OpenClaw / QwenPaw / Hermes Agent (264 GitHub starsA lightweight x86-64 / arm64 Virtual Machine Monitor (VMM) for OpenClaw / QwenPaw / Hermes Agent (264 GitHub stars; NOASSERTION; owner 78) |
| Local | Local | Free, open source (MIT); runs on your own MacPricing page | N/a (runs locally) | VM persists and is reused on later runs; pause/resume and stop (normal shutdown orVM persists and is reused on later runs; pause/resume and stop (normal shutdown or --force); VM cloning and rollback checkpoints listed as future work | VM | Nov 2026 | Separate macOS desktop for computer-use agents, viewed inline in Ghostty, Kitty or iTerm2 or in a native windowSeparate macOS desktop for computer-use agents, viewed inline in Ghostty, Kitty or iTerm2 or in a native window; MCP tools screenshot, click, drag, scroll, type, key, status; Claude Code plugin and skill; CLI actions return JSON; multiple VMs; default disk 64 GiB sparse; guest setup (account, Screen Recording and Accessibility permissions) is manual; Apple Silicon, macOS 14+; self-signed releases; 2,737 GitHub stars |
| Local | Local | Free, open source (no license file)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2025 | Run Claude Code safely in local Docker containers without having to approve every permission (ArchivedRun Claude Code safely in local Docker containers without having to approve every permission (Archived, see Spritz repo for the continuation of this project) (324 GitHub stars; no license file; archived; owner textcortex) |
| Local | Local | Free open source (MIT); runs on your own machine | — | Ephemeral sandbox per run; daemon mode reuses one sandbox across projectsEphemeral sandbox per run; daemon mode reuses one sandbox across projects; --clone runs disposably and exports a git bundle patch | VM | Dec 2025 | Single-binary CLI that starts a sandbox preloaded with 14 coding agents (Claude Code, Codex, Copilot, OpenCode, AmpSingle-binary CLI that starts a sandbox preloaded with 14 coding agents (Claude Code, Codex, Copilot, OpenCode, Amp, Goose and others); macOS, Linux, Windows (WSL); host-side credential proxy keeps real API keys out of the sandbox; network policy tiers (allowlist, open, internet, none); locks host .git; SSH agent forwarding; clipboard bridge; MIT, 61 GitHub stars |
| Local | Local | Free, open source (MIT); runs on your own Mac (Apple silicon, macOS 14+)Free, open source (MIT); runs on your own Mac (Apple silicon, macOS 14+); built from source | N/a (runs locally) | Disposable sandboxes; pause to disk and resume (~0.5-0.7 s)Disposable sandboxes; pause to disk and resume (~0.5-0.7 s); stopping the foreground daemon pauses sandboxes to disk; checkpoint, restore and fork snapshots; warm pool | VM | Oct 2026 | Local daemon (sandboxd) runs one microVM per agent; Cursor, Claude Code and Codex connect over MCP (sandbox-mcp)Local daemon (sandboxd) runs one microVM per agent; Cursor, Claude Code and Codex connect over MCP (sandbox-mcp), or code calls the HTTP API on a unix socket; repo synced into the VM; egress through an allowlist proxy, secrets never enter the VM; claims: warm-pool create 20-50 ms, snapshot restore ~0.5 s, cold boot ~1.2 s (M4 MacBook); vCPU/memory scheduler with memory-pressure handling; no Windows; 134 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | May 2025 | Keeping your agents out of trouble with sandboxed coding agent automation (172 GitHub stars; MIT; owner dtormoen) |
| Local | Local | Free, open source (MIT)Pricing page | N/a (self-hosted) | SDK has reconnect and kill; no idle policy stated | Container | Jun 2026 | E2B-style TypeScript SDK; runs on the user's own Mac (macOS 26+, Apple container CLI), no Docker daemonE2B-style TypeScript SDK; runs on the user's own Mac (macOS 26+, Apple container CLI), no Docker daemon; any Linux OCI image; CLI, HTTP API (Hono) and MCP server; LangChain deepagents and Mastra backends; Node or Bun; early development, pre-1.0; 82 stars; all pushes on 2026-06-23 |
| Local | Local | Free, open source (Apache-2.0)Pricing page | N/a (runs on your own machine) | No daemon; the sandbox lasts as long as the wrapped agentNo daemon; the sandbox lasts as long as the wrapped agent; a copy-on-write workspace snapshot is taken before each session (vetto diff, vetto undo) | Process | Aug 2026 | Rust CLI via install script, npm (@shledery/vetto), Homebrew or cargoRust CLI via install script, npm (@shledery/vetto), Homebrew or cargo; prebuilt for Linux x86_64/aarch64, macOS Apple silicon/Intel and Windows x86_64; vetto enable --all wraps detected agents (claude, codex, cursor, opencode, aider) with shims; agent compatibility registry of 32+ agents; host credentials masked and network scoped to the agent's API; startup overhead under 4 ms on Linux, 8 ms on macOS, 15 ms on Windows; vetto doctor and vetto verify; Merkle audit log per session; GitHub Actions integration; the linked GitHub repo and account return 404 |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Easy Linux virtual machine on MacOS to sandbox LLM agents. (963 GitHub stars; MIT; owner lynaghk) |
| Local | Local | Free, open source (NOASSERTION)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Jan 2026 | Vibebin: code and host inside Incus containers on your own VPS/server. (111 GitHub stars; NOASSERTION; owner jgbrwn) |
| Local | Local | Free, open source (MIT); runs on your own MacPricing page | N/a (runs locally) | Per-project instance disk persists in .vibebox/ and is reusedPer-project instance disk persists in .vibebox/ and is reused; default config sets supervisor auto_shutdown_ms = 20000 (the README doesn't say what triggers it); vibebox reset deletes the instance | VM | Feb 2026 | Per-repo sandbox entered with one command; warm re-entry typically under 5 s on an M3 (author's figure)Per-repo sandbox entered with one command; warm re-entry typically under 5 s on an M3 (author's figure); default 2 CPUs, 2048 MB RAM, 5 GB disk, set in vibebox.toml; .git masked with tmpfs inside the VM; Codex and Claude Code installed on first login (best effort); multiple terminals and sessions; crates.io package; Apple Silicon only; site vibebox.robcholz.com returns HTTP 525; 196 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | N/a (runs on your own machine) | Disposable: each run goes INIT, PROVISIONED, STAGED, SEALED, RUNNING, CAPTUREDDisposable: each run goes INIT, PROVISIONED, STAGED, SEALED, RUNNING, CAPTURED, EXTRACTED, DESTROYED, and teardown always runs (VM removed, created disks deleted) | VM | Jun 2026 | PowerShell 7 engine; Windows 11 Pro with Hyper-V is the tested hostPowerShell 7 engine; Windows 11 Pro with Hyper-V is the tested host; a host-verified, fail-closed seal gate checks that the VM has no network, credentials or host-guest channels beyond its tier before the workload runs; inputs and outputs travel on attached data disks with no live channel; calls itself "Claude-Code-native" and the "bring your own VM" step above Claude Code's process sandbox; the README says only the Tier-0 disk round trip has run live and the rest is mock-tested; by Polycentric Labs, LLC; 1 GitHub star |
| Local | Local | Free, open source, runs on your own machine (the repo has no license file)Pricing page | N/a (runs locally) | One persistent machine per project directory: ./wrapvm creates or starts it and opens aOne persistent machine per project directory: ./wrapvm creates or starts it and opens a shell, with the directory mounted as the workspace; no auto-suspend or timeout stated | VM | Nov 2026 | Desktop VM image for agents: Xvfb + XFCE viewed in a browser through voxl on port 6080Desktop VM image for agents: Xvfb + XFCE viewed in a browser through voxl on port 6080; one persistent Google Chrome with CDP on 9222; Vercel Labs' agent-browser and agent-desktop CLIs plus desktop and browser skills for agents; PulseAudio audio, streamed and recordable; ffmpeg; Smolfile defaults to 4 vCPU / 8 GiB and installs pi, codex and claude on first start; network blocks the LAN, host loopback, cloud metadata and CGNAT ranges; the host injects GitHub credentials as a request header so the token stays off the guest; image ghcr.io/tobi/wrapvm:desktop, built on tobi/wrap; by Tobias Lütke; 382 GitHub stars |
| Local | Local | Free, open source (the README says MIT; GitHub detects the license as "Other")Pricing page | N/a (runs on your own machine) | Sandboxes persist per branch until yolo-cage deleteSandboxes persist per branch until yolo-cage delete; yolo-cage up/down start and stop the VM | VM | Jan 2026 | Runs Claude Code in YOLO mode; the agent can push only to its assigned branch; TruffleHog pre-push scanningRuns Claude Code in YOLO mode; the agent can push only to its assigned branch; TruffleHog pre-push scanning; blocks gh pr merge, gh repo delete, gh api and GitHub merge/delete API calls; scans HTTP traffic for sk-ant-, AKIA and ghp_ keys and SSH private keys; blocks paste sites; port forwarding; needs 8 GB RAM, 4 CPUs, a GitHub PAT and a Claude account; the README lists DNS exfiltration and steganography as unhandled; 109 GitHub stars |
| Local | Local | Free, open source (MIT)Pricing page | $0 (self-hosted) | Runs on your own machine or servers | — | Feb 2026 | Your agent is a security risk, so treat it like one. yoloAI does AI agent sandboxing right. (256 GitHub stars; MITYour agent is a security risk, so treat it like one. yoloAI does AI agent sandboxing right. (256 GitHub stars; MIT; owner kstenerud) |
| Local | Local | Free, open source (MIT)Pricing page | N/a (runs on your own machine) | A container per run; named persistent volumes keep tools, configsA container per run; named persistent volumes keep tools, configs, caches and agent sessions across runs and are shared by boxes on the same runtime; --scratch runs without them; yolobox reset removes the volumes | Container | Jan 2026 | Go CLI via Homebrew or install script; shortcuts for Claude Code, Codex, Kimi Code, Gemini, Antigravity, OpenCodeGo CLI via Homebrew or install script; shortcuts for Claude Code, Codex, Kimi Code, Gemini, Antigravity, OpenCode, Copilot and Pi, with permission prompts skipped inside the box; base image with Node.js, Python, Go, Bun, Git, GitHub CLI, ripgrep and more, and the agent has sudo; per-project .yolobox.toml for packages and Dockerfile fragments; --no-network, --readonly-project, file excludes and host env aliasing for read-only tokens; yolobox fork gives an agent its own project copy; Codex and Claude ACP adapters for editors; by Finbarr Taylor; Show HN 2026-01-12 (122 points); 657 GitHub stars |
| Local | Local | Free, open source (Apache-2.0); runs on your own machinePricing page | N/a (runs locally) | Per command, no persistent sandbox; --snapshot records filesystem changes and --restorePer command, no persistent sandbox; --snapshot records filesystem changes and --restore undoes them after the run | Process | Mar 2026 | Deny by default: writes, network and environment variables blocked unless allowed; per-path read/write rulesDeny by default: writes, network and environment variables blocked unless allowed; per-path read/write rules; outbound filtering by domain; secrets injected at a proxy only for approved hosts, so the process sees placeholders; SDKs for Rust, TypeScript and Python; single binary, about 10 ms overhead; npm, PyPI, crates.io; 730 GitHub stars |
| Platform feature | Platform feature | Not published; request accessPricing page | — | Site mentions durable tasks, hot snapshots and recovery | — | — | Agent swarms: one CLI command (@acyclic-labs/cli) splits a task across many agents, each in its own sandboxAgent swarms: one CLI command (@acyclic-labs/cli) splits a task across many agents, each in its own sandbox, and merges the output into one patch or pull request; dashboard and MCP; BYOK model keys; BYOC sandboxes in your AWS, GCP or Azure account; site says the harness is open source |
| Platform feature | Platform feature | Credits at $0.10 each in packs ($10 per 100)Credits at $0.10 each in packs ($10 per 100), metered per second of active runtime: Medium 2 vCPU/8 GB 10 credits/h ($1.00/h), Large 4 vCPU/16 GB 20/h ($2.00/h), XL 8 vCPU/32 GB 40/h ($4.00/h); the homepage also lists a default Small 1 vCPU/4 GB at 5 credits/h ($0.50/h), which the docs omit; no 2 vCPU/4 GB size; plans Free $0 (20 credits/mo), Starter $30/mo (400), Pro $75/mo (2,000), Max $120/mo (4,000), Teams $100/seat/mo (2,500 per seat), Enterprise custom; Android emulator 60 credits/h, iOS simulator 120 credits/h; GB treated as GiB | $0 (idle and suspended devboxes bill nothing) | Workspace suspends when idle (memory and disk snapshot) and resumes where it left offWorkspace suspends when idle (memory and disk snapshot) and resumes where it left off; keep-active opt-out per workspace; machine and volume destroyed when the task's PR merges or closes, or after 48 hours of inactivity | VM | Feb 2026 | Runs Claude Code, Codex or OpenCode on your own subscription or API key (40+ models)Runs Claude Code, Codex or OpenCode on your own subscription or API key (40+ models); a warm slot starts a task in about 5 s; Ubuntu 22.04 devbox with Docker, Playwright, Go, Rust, Node, Bun, Python; terminal, file editor, port previews and SSH (port 2222) into the devbox; automated PR reviews with screenshots and video; campaigns split work across sub-agents in parallel devboxes; REST and WebSocket API, CLI, MCP, Slack, Linear, GitHub, Sentry, mobile; Android emulators and iOS simulators; macOS app; one region (San Jose); no SOC 2 yet; Enterprise adds self-hosting and BYOC |
| Platform feature | Platform feature | Software free for now (no licence, key or charge; AGPL-3.0)Software free for now (no licence, key or charge; AGPL-3.0); you pay Fly directly: the installer estimates about $17/month (about $0.57/day) for the Workspace, code host and router, plus about $0.09/hour for each awake agent at Fly's list prices; email about $4/month; Claude, ChatGPT or OpenCode plans billed by the AI provider (price per hour for 2 vCPU / 4 GiB: n/a (self-hosted; Fly bills you directly, about $0.09/h per awake 2-core / 4 GB agent by the site's estimate)) | No charge from the vendor; Fly bills the sleeping agent's disk (amount notNo charge from the vendor; Fly bills the sleeping agent's disk (amount not stated) | Each agent's machine sleeps about 3 quiet minutes after its last turn and wakes inEach agent's machine sleeps about 3 quiet minutes after its last turn and wakes in seconds on a message or when a page it serves is opened (about 10 s first load); its disk is kept while it sleeps and grows as it fills; containers inside stop on sleep; deleting an agent erases its machine | VM | Jul 2026 | Self-hosted workplace for a team of Claude Code, Codex and OpenCode agentsSelf-hosted workplace for a team of Claude Code, Codex and OpenCode agents, each with its own machine (default 2 dedicated cores and 4 GB, resizable); a 'Chief of Staff' agent hands out jobs; built-in git hosting with CI/CD, previews and site hosting on your domain, object storage, email, Drive, notes and knowledge pages; changes to protected code wait for human approval unless auto-merge is on; one-command install (13 min 24 s in a 2026-10-29 run) and upgrades; sends the vendor a daily count report that the admin can turn off; read-only GitHub mirror adambaialiev/agentic-cloud (0 stars); the Show HN version was a Telegram bot |
| Platform feature | Platform feature | Self-hosted open source; no hosted service or prices | — | Each request spawns an agent session that runs to completion and is destroyedEach request spawns an agent session that runs to completion and is destroyed; fixed pool of pre-built sandbox slots; user workspace and skills persist across sandbox teardown | Container | Aug 2026 | Self-hosted multi-tenant agent service built on the pi coding agent; Vue 3 chat UI with SSE streamingSelf-hosted multi-tenant agent service built on the pi coding agent; Vue 3 chat UI with SSE streaming; versioned artifacts with share links and a public market; long-term memory (MEMORY.md); projects; cron scheduled tasks; skills staged into the sandbox; Docker Compose (agent, worker, manager images) with MySQL and Redis; no LICENSE file; 27 GitHub stars |
| Platform feature | Platform feature | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (Apache-2.0) | — | — | Container | Dec 2025 | Self-hosted AI coding workspace that runs Antigravity, Claude Code, Codex, Copilot, CursorSelf-hosted AI coding workspace that runs Antigravity, Claude Code, Codex, Copilot, Cursor, Grok and OpenCode through ACP adapters; each workspace gets its own Docker or host sandbox; chat, editor, terminal, diffs, secrets and git tools; multi-agent orchestration through a bundled MCP server; group 'channels' of agents; ships as a Docker web app, a macOS desktop app and an iOS app; Apache-2.0; 339 stars |
| Platform feature | Platform feature | Open-source framework feature; backend costs (E2B, Daytona) are separate | — | Set by the backend: Docker and bubblewrap are ephemeral unless a host workdir is mountedSet by the backend: Docker and bubblewrap are ephemeral unless a host workdir is mounted; E2B, Daytona, OpenSandbox and K8s workspaces are reattached by metadata, labels or PVC | Container | May 2026 | Execution-environment layer of the AgentScope 2.0 agent framework: built-in tools (Bash, ReadExecution-environment layer of the AgentScope 2.0 agent framework: built-in tools (Bash, Read, Write) run through the workspace backend, MCP servers, skills and context offloading; seven implementations behind one interface; AgentScope is Apache-2.0 with 36k GitHub stars |
| Platform feature | Platform feature | Free $0/month (evaluation: 1 workspace, 2 agents at 1 vCPU / 1 GB RAM eachFree $0/month (evaluation: 1 workspace, 2 agents at 1 vCPU / 1 GB RAM each, shared multi-tenant SaaS, 7-day audit retention; the FAQ says up to 3 members while the plan table says unlimited users); Enterprise custom pricing (SaaS, on-prem, self-deployed or air-gapped); bring your own model key or subscription | — | Agents are stateful: the home directory is on a Kubernetes persistent volume and survivesAgents are stateful: the home directory is on a Kubernetes persistent volume and survives restarts; workflow timeout and retention are editable; no idle or suspend policy stated | Container | Apr 2026 | Zero-trust agent platform; default-deny egress with per-sandbox allowed hostsZero-trust agent platform; default-deny egress with per-sandbox allowed hosts; secret-injection proxy, so agents see placeholders and never the credentials; reusable sandbox configs of models, skills, MCP tools, software packages and allowed hosts; workflows on a schedule or webhook; OpenCode TUI can attach to an agent; inference through Bedrock, Vertex, Azure AI Foundry, OpenAI or custom endpoints; open source (Apache-2.0, 56 GitHub stars), built on OpenCode, AgentGateway, OpenBao, Cilium and KubeArmor |
| Platform feature | Platform feature | CNY only, pay as you go in CUs: compute (runtime) CNY 0.0000694 per CU in BeijingCNY only, pay as you go in CUs: compute (runtime) CNY 0.0000694 per CU in Beijing, Hangzhou, Shanghai and Shenzhen, where 1 vCPU-second = 1 CU and 1 GB-second = 0.5 CU (CNY 0.2498/vCPU-h + 0.1249/GB-h; 2 vCPU / 4 GB about CNY 0.9994/h); Singapore and Hong Kong CNY 0.0001 per CU (about CNY 1.44/h for 2 vCPU / 4 GB); gateway CNY 0.0000015 per call-CU; storage CNY 0.00025 per CU (10 GB for 720 h = CNY 1.8); optional dedicated control plane CNY 3-4.5 per hour; internet traffic CNY 0.8/GB. No USD price on the site, so the USD hourly figure is '?' | — | — | — | — | Agent building and governance platform: agents configured with SOUL.MD and AGENT.MD, sub-agents, teamsAgent building and governance platform: agents configured with SOUL.MD and AGENT.MD, sub-agents, teams, model connections, a skill marketplace, MCP tools, credential vault, memory and knowledge bases, observability and evaluation; agents reachable through DingTalk, WeChat and Feishu; regions Hangzhou, Shanghai, Shenzhen, Beijing and Singapore; depends on CMS, PrivateLink and SchedulerX3; the sandbox runtime only runs agents built in AgentCore; no standalone sandbox API described |
| Platform feature | Platform feature | $0.17 | $0 (a paused orb costs nothing) | Auto-pauses when the agent has done no work for 5 min and nobody has interacted for 20Auto-pauses when the agent has done no work for 5 min and nobody has interacted for 20 min; a paused orb keeps the conversation, files and running services and wakes on the next message, terminal/desktop use or portal request; archiving a thread pauses its orb at once | — | Jun 2026 | Cloud machine per Amp thread with the repo, plugins and tools; Files, Terminal and Desktop panesCloud machine per Amp thread with the repo, plugins and tools; Files, Terminal and Desktop panes; portals (shareable URLs to services); agents spawn other agents in fresh orbs; amp sync to pull changes locally; self-hosted orbs and Mac satellites; spawn pacing of 20 orbs then one every 5 min; works only inside Amp |
| Platform feature | Platform feature | VPS + AI credits plans: Free $1/mo (2 vCPU, 2GB, 20GB, 5,000 credits)VPS + AI credits plans: Free $1/mo (2 vCPU, 2GB, 20GB, 5,000 credits); Pro $39/mo (4 vCPU, 8GB, 40GB, 20,000 credits); Ultra $79/mo (8 vCPU, 16GB, 80GB, 40,000 credits); Business custom (16 vCPU, 32GB, 200GB); $1 = 1,000 credits, refreshed monthly; FAQ says free hosting during beta with $10 AI credits; a 'VPS Only' price tab exists but its prices are not in the page HTML; no 2 vCPU / 4 GiB plan | N/a (always on, flat monthly plans) | Always on ('Just your OpenClaw agent, always on'); no sleep or idle policy stated | — | Feb 2026 | Managed OpenClaw hosting ('Create and setup your bot in seconds', no server to manage)Managed OpenClaw hosting ('Create and setup your bot in seconds', no server to manage); BYOK with zero markup on Pro and up; unlimited web search, browser automation, cloud backup, auto update; integrations such as Slack, Notion, GitHub, Stripe, Gmail; one-click export; import from a self-hosted OpenClaw in under 5 minutes |
| Platform feature | Platform feature | Managed Agents: tokens at normal model rates + $0.08 per session-hour of running timeManaged Agents: tokens at normal model rates + $0.08 per session-hour of running time (measured to the millisecond), no separate container-hour charge; web search $10 per 1,000; no CPU/RAM shape published (sandbox: up to 8 GB memory, 10 GB disk). Claude Code cloud sessions: included in Pro, Max, Ultra, Team and eligible Enterprise plans, counted against plan usage limits, no separate charge for the cloud VM. Messages API code execution tool: $0.05/container-hour after 1,550 free hours per org per month (5-minute minimum) | $0 (idle session time not billed) | Managed Agents: session runtime accrues only while status is runningManaged Agents: session runtime accrues only while status is running; idle time (waiting for your next message or a tool confirmation), rescheduling and terminated time are not billed. Claude Code cloud: an idle session can be resumed until its environment expires | Container | Apr 2026 | Managed Agents: managed agent harness for Claude only (agents, environments, sessions, SSE event stream)Managed Agents: managed agent harness for Claude only (agents, environments, sessions, SSE event stream), Ubuntu 24.04 x86_64 sandbox with Python, Node, Go, Rust, Java, Ruby, PHP, C/C++, PostgreSQL, Redis, Playwright + Chromium, LibreOffice, TeX; network unrestricted by default for API-created environments; self-hosted sandboxes or a fully self-hosted service also available; memory, webhooks, vaults, multi-agent orchestration. Claude Code cloud: runs Claude Code sessions from browser, phone, desktop or terminal (--cloud, --teleport), GitHub clone and push, auto-fix PRs |
| Platform feature | Platform feature | Free to try with 500 credits, no card; credits cover what agents spend on modelsFree to try with 500 credits, no card; credits cover what agents spend on models; pay-as-you-go after that, with no published rates; SSO, self-hosting or custom SLAs by call | — | Sessions snapshot their filesystem and chat state automatically when idleSessions snapshot their filesystem and chat state automatically when idle (idle_timeout_s) and show as paused; resume() restores from any snapshot | VM | May 2026 | 'Platform for every agent your team builds': define an agent once (repo, install command, skills, MCP servers)'Platform for every agent your team builds': define an agent once (repo, install command, skills, MCP servers), AnyFrame bakes a cached image, and each session boots a fresh Ubuntu sandbox (the demo shows 4 vCPU) running Claude Code, Codex, Cursor, OpenCode or Gemini CLI; triggers from Slack, Discord, Linear, GitHub and Jira; lazy VNC desktop sandboxes and a browser_use tool; Python SDK (MIT, 16 stars) and TypeScript SDK; MCP connector marketplace; open-source version 'coming soon'; moved from anyfrm.com to anyframe.dev in May 2026 |
| Platform feature | Platform feature | Pricing page not readable: the site answers every request with a Vercel SecurityPricing page not readable: the site answers every request with a Vercel Security Checkpoint, HTTP 429 | — | SDK: computers are started, listed and deleted through the API | VM | May 2025 | From the first-party Python SDK (github.com/ms337/autocomputer-sdk, author madhav@autocomputer.ai, MITFrom the first-party Python SDK (github.com/ms337/autocomputer-sdk, author madhav@autocomputer.ai, MIT, v0.9.0 beta): provisions remote Linux computers through api.autocomputer.ai, uploads and downloads files, gives a VNC URL for live viewing, and runs JSON-defined natural-language workflows with AutoComputer's own computer-use agent, streaming progress; can also run workflows on a local VirtualBox VM |
| Platform feature | Platform feature | Machine time per hour while awake, plus a 2-minute wind-down at each sleep: Small 1Machine time per hour while awake, plus a 2-minute wind-down at each sleep: Small 1 vCPU/4 GB $0.10; Medium 2 vCPU/8 GB $0.20; Large (default) 4 vCPU/16 GB $0.40; Ultra 8/32 $0.80; Hyper 16/64 $1.60; Big Guy 16/128 $3.20 (on request). Plans are paid as credits: Lite $20/mo ($20 credits), Pro $100/mo ($105), Max $200-$1,000/mo (110% in credits), annual 20% off; tokens on your own subscriptions or keys are free, machine time always uses credits. No 2 vCPU/4 GB size: nearest are Small 1 vCPU/4 GB $0.10/h and Medium 2 vCPU/8 GB $0.20/h | $0 (a sleeping machine costs nothing); 2-minute wind-down billed at each sleep | A cloud machine sleeps about 2 minutes after the thread stops workingA cloud machine sleeps about 2 minutes after the thread stops working, unless a background command (up to 6 h), an open shell/desktop/preview tab, preview traffic or "Keep awake 4h" holds it; the disk survives sleep, memory and processes do not; a machine asleep for 30 days is deleted | VM | 2024 | Hosted coding agent (Captain orchestrates, Build writes codeHosted coding agent (Captain orchestrates, Build writes code, Review agent posts PR findings) that opens and fixes pull requests; each cloud thread runs on its own Linux VM: Ubuntu 24.04 x86_64, passwordless sudo, Docker with Compose, XFCE desktop with Google Chrome streamed to the thread pane, no /dev/kvm; region US East (New York); project snapshots restore machines in seconds; iOS and Android cloud phones on Pro, Max and Enterprise; Tailscale tailnet join; threads start from web or desktop app, Slack, Linear or API; local threads on your own computer through the desktop app; running-memory caps Lite 128 GB, Pro 512 GB, Max 1 TB; SOC 2 Type II |
| Platform feature | Platform feature | Self-hosted open source; no hosted service or prices | — | Sandbox paused after an idle timeout (Slack bot default up to 3 hours)Sandbox paused after an idle timeout (Slack bot default up to 3 hours); every sandbox, running or suspended, deleted after sandboxMaxLifetimeSecs (default 72 hours); warm pool | Container | May 2026 | Self-hosted Slack agent platform for teams: each Slack thread gets its own sandbox with shell, git, Python, Node.jsSelf-hosted Slack agent platform for teams: each Slack thread gets its own sandbox with shell, git, Python, Node.js, Bun; bring your own harness (Amp, Claude Code, Codex); credentials injected at the network edge by iron-proxy; durable workflows; HTTP API; MIT licensed (site); 1.4k GitHub stars |
| Platform feature | Platform feature | Starter $39/month (1 private agent workspace, 1,000 managed provider credits/month)Starter $39/month (1 private agent workspace, 1,000 managed provider credits/month); Pro $99/month (1 higher-limit workspace, browser automation and code execution, API access, 2,000 credits); Done-For-You $1,500 (setup, 3 months of Pro, $300 provider credits); BYOK available; no CPU or RAM published | Same as running (flat monthly plan) | Always on (24/7 managed OpenClaw server); lifecycle controls mentioned in the changelogAlways on (24/7 managed OpenClaw server); lifecycle controls mentioned in the changelog, no suspend described | VM | Jun 2026 | Managed OpenClaw install on its own cloud server: server provisioning, Telegram bot pairing tested before deployManaged OpenClaw install on its own cloud server: server provisioning, Telegram bot pairing tested before deploy, managed model routing, protected dashboard; premium skills; Cloud Pilot program for selected users; a 'reviewable business workbench' is described as future direction; product of Breydan Tech Solution |
| Platform feature | Platform feature | $0.0137 | Same as running (stopping doesn't cancel the monthly compute subscription) | Cloud Agents are always on; Stop pauses the remote runtime (browser tools and TerminalCloud Agents are always on; Stop pauses the remote runtime (browser tools and Terminal unavailable until the next start) and doesn't cancel the compute subscription; Delete permanently removes the agent and releases its resources | — | Feb 2026 | Hosted Hermes Agent or OpenClaw runtime per Cloud Agent, reached over Telegram, Discord or WhatsAppHosted Hermes Agent or OpenClaw runtime per Cloud Agent, reached over Telegram, Discord or WhatsApp; browser interface, workspace files and Terminal; memories, skills, vaults and OAuth connectors shared with local agents (Claude Code, Codex, Hermes, OpenClaw, Pi, OpenCode) through the clawdi CLI; CLI, backend and dashboard open source (MIT, GitHub Clawdi-AI/clawdi 107 stars) and self-hostable with PostgreSQL; operated by Goodwill Labs Inc. |
| Platform feature | Platform feature | Early access; no price published. Umans AI's main product is pay-per-token inference forEarly access; no price published. Umans AI's main product is pay-per-token inference for open models | — | — | — | — | 'Cloud agents is the next chapter of envs': connect GitHub or GitLab and an agent plans, edits'Cloud agents is the next chapter of envs': connect GitHub or GitLab and an agent plans, edits, runs your tests and full builds in a real environment, driven from the browser; from Umans AI, a European inference provider for open coding models |
| Platform feature | Platform feature | Hard-capped plans with hosted models (no API keys): Free $0 (50 AI creditsHard-capped plans with hosted models (no API keys): Free $0 (50 AI credits, 10 cloud browser minutes, 10 web searches a month); Growth $29/mo (2,500 credits, 500 browser minutes, 120 searches); Pro $59/mo (5,000 credits, 1,000 minutes, 240 searches); Max $199/mo (18,000 credits, 3,500 minutes, 800 searches); $5 starter pack of 100 bonus credits; list prices since 2026-09-01, founding members keep their price through 2027-03-01; no compute size stated | Included in plan | Persistent cloud desktop that keeps running when you close the tab: filesPersistent cloud desktop that keeps running when you close the tab: files, browser sessions, logins and schedules stay alive 24/7 | — | 2025 | Cloudia, CloudAxis's own agent, works on a persistent cloud desktop with a cloud browser routed from your countryCloudia, CloudAxis's own agent, works on a persistent cloud desktop with a cloud browser routed from your country, Word, Excel and PDF apps, image and video generation, posting to 13 social platforms, and Gmail, Outlook, Shopify, Notion and GitHub integrations; scheduled jobs and a multi-agent team board; Auto Mode picks the model (GPT-5.6 Luna), with no model picker; the same company runs CloudyBot |
| Platform feature | Platform feature | Free, open source (Apache-2.0) to self-hostFree, open source (Apache-2.0) to self-host; you supply and pay for Daytona, Convex, Clerk and Trigger.dev yourself; the hosted app at cloudcode.computer shows no pricing | — | Creates each Daytona sandbox with auto-stop after a configurable number of idle minutesCreates each Daytona sandbox with auto-stop after a configurable number of idle minutes and restarts a stopped sandbox on the next run (source code; not documented on the site) | — | May 2026 | Web app for chatting with Codex in a Daytona sandbox; sandbox comes with a desktop and terminalWeb app for chatting with Codex in a Daytona sandbox; sandbox comes with a desktop and terminal; Codex records videos of UI changes it tested on the desktop; MCP servers; shared project notes for agents and humans; SSH access; GitHub App for PR reviews and automation triggers; Apache-2.0, 15 GitHub stars |
| Platform feature | Platform feature | Free $0 with 3,000 Cloud Credits/month; Pro $50/mo with 50,000Free $0 with 3,000 Cloud Credits/month; Pro $50/mo with 50,000; Super $150/mo with 150,000; Ultra $500/mo with 500,000; yearly 17% off; a typical hour of agent work uses about 100 credits (light work about 20), counted only while the agent actively works; AI usage stays on your own plan | $0 (idle and sleeping sandboxes and storage use no credits) | Sandbox sleeps after 5 idle min (Free), 20 (Pro), 30 (Business)Sandbox sleeps after 5 idle min (Free), 20 (Pro), 30 (Business), 40 (Super) or 60 (Ultra); never while the agent works; next message wakes it with files and history; archived thread's sandbox saved and removed after 2 h; unused threads deleted after the plan's storage limit (7 days Free to 5 years Ultra) | VM | Sep 2026 | macOS desktop app that runs coding agents locally or in cloud sandboxes: Claude Code, Codex, Pi, OpenCode, fx, HermesmacOS desktop app that runs coding agents locally or in cloud sandboxes: Claude Code, Codex, Pi, OpenCode, fx, Hermes, Devin, Droid, Kimi Code; uses your Claude, ChatGPT, Cursor or Copilot plans or API keys; own Git branch per thread; Teleport moves live sessions between Mac and cloud; H100 and B200 GPUs on demand on paid plans; full admin inside each sandbox; Apache-2.0, 1,666 GitHub stars |
| Platform feature | Platform feature | Plans with usage allowances and optional pay-as-you-go: Free $0 (CPU sandbox with 8 vCPUPlans with usage allowances and optional pay-as-you-go: Free $0 (CPU sandbox with 8 vCPU / 8 GB RAM, 5 GB checkpoint storage, 1 concurrent sandbox); Plus $12/mo (T4 GPU, 16 GB RAM, 30 GB, 3 concurrent); Pro $30/mo (T4, L4 or A10, 32 GB RAM, 100 GB, 8 concurrent); Max $90/mo (10x Plus usage, up to A100, 64 GB RAM, 512 GB, 12 concurrent) or $200/mo (30x, up to H200, 128 GB RAM, 2 TB, 24 concurrent); model access widens with the plan; Enterprise by call | Paused sandboxes are covered by the plan; no separate rate | On paid plans an idle sandbox is paused in place and resumes with its memoryOn paid plans an idle sandbox is paused in place and resumes with its memory; on Free it is released after 3 days, while the project and its saved kernel state (checkpoint storage) stay restorable | — | Jun 2026 | AI notebook IDE for Jupyter: Clusy's own agent plans the work and writes and runs notebook cells on cloud CPUs or GPUsAI notebook IDE for Jupyter: Clusy's own agent plans the work and writes and runs notebook cells on cloud CPUs or GPUs; parallel experiments on separate branches, each with its own kernel; snapshot a workspace and move it to a bigger GPU, then resume from the checkpoint; data connectors for Hugging Face, Kaggle, GitHub, Snowflake, Databricks, S3, BigQuery and others; models from an Auto model up to Claude and GPT on Max. Clusy also sells Portal (portal.clusy.io), which moves running Python jobs between clouds and CPU/GPU architectures from inside the process so they can use spot compute |
| Platform feature | Platform feature | Plans per workspace: Free $0 (200 credits a month, Starter and Small computers)Plans per workspace: Free $0 (200 credits a month, Starter and Small computers); Plus $30/mo (2,500 credits); Pro $100/mo (10,000 credits, all profiles); Enterprise custom. Compute profiles in credits per running hour, prorated by the second: Starter 1 vCPU / 4 GB 7.5; Small 2 vCPU / 4 GB 10; Standard 2 vCPU / 8 GB 15; Compute 4 vCPU / 8 GB 22; Large 4 vCPU / 16 GB 30; X-Large 8 vCPU / 32 GB 60. Snapshot storage 8 (Free), 6 (Plus, Pro) or 4 (Enterprise) credits per GiB-month. Extra credits can be bought, but no per-credit price is published (Plus's included credits work out to $0.012 each) | $0 compute while suspended; retained snapshots 4-8 credits per GiB-month | Suspend stops compute charges; optional auto-suspend after the agent's response (off bySuspend stops compute charges; optional auto-suspend after the agent's response (off by default); docs mention 'ordinary idle suspension' without a stated timeout; tasks resume later; a computer can be saved as a reusable starting point; retained snapshots keep billing storage | — | Jul 2026 | Cloud harness that gives each coding-agent task its own computer with repo checkout, servicesCloud harness that gives each coding-agent task its own computer with repo checkout, services, live preview and pull request; agents: Codex, Cursor, Claude Code, GitHub Copilot, Gemini, Grok, Muse Code, OpenCode, Pi (bring your own accounts); Loops for scheduled work; Agent Browser to test the running app; team collaboration; Vault for secrets; GitHub, Azure DevOps and GitLab; Cobalt Desktop app (2026-11-05); bring-your-own compute 'coming soon'; the same company sells Borealis AI Sandboxes (borealishq.io) |
| Platform feature | Platform feature | Not published (/pricing redirects to 'schedule a demo')Not published (/pricing redirects to 'schedule a demo'); Enterprise plans add hard daily and monthly spend caps | $0 compute while paused (per docs); storage price not published | Sleep pauses the computer by default: files, processes, browser sessions and loginsSleep pauses the computer by default: files, processes, browser sessions and logins, installed packages and memory are frozen, resume takes about 1 s, and a paused computer has no compute cost; on-demand mode destroys the computer on sleep and starts a fresh one on wake; a Telegram message wakes a sleeping agent | VM | — | AI operations agents for business workflows (invoices, reconciliation, vendor follow-up, reports)AI operations agents for business workflows (invoices, reconciliation, vendor follow-up, reports), each with its own computer for real computer use across browser, apps, files and terminal; live desktop preview; recordable, replayable action sequences; channels: email inbox, WhatsApp, Telegram, phone calls in 57+ languages, webhooks and API; role-aware memory and permissions for teammates and outside people; credential vault; PII redaction (Enterprise); Communa also builds managed agents for customers |
| Platform feature | Platform feature | Sandboxes are not billed today; Composio says it plans to bill by tier and runtime soonSandboxes are not billed today; Composio says it plans to bill by tier and runtime soon (no rates published); tiers: standard 1 vCPU/1 GB (default), medium 2 vCPU/2 GB, large 4 vCPU/4 GB, xlarge 8 vCPU/8 GB, so no 2 vCPU / 4 GB tier; Composio plans: Free $0 ($2 one-time credits), Standard $18/seat/mo ($18 credits per seat), extra tool calls $0.005 | Not billed today | Persistent Jupyter-notebook Python environment per session: variablesPersistent Jupyter-notebook Python environment per session: variables, imports and files persist across calls within a session; /mnt/files mount survives sandbox restarts; changing the compute tier recreates the sandbox and clears in-memory state; no idle timeout stated | — | — | Formerly called the workbench; agents run Python via the COMPOSIO_REMOTE_WORKBENCH meta tool and shell viaFormerly called the workbench; agents run Python via the COMPOSIO_REMOTE_WORKBENCH meta tool and shell via COMPOSIO_REMOTE_BASH_TOOL inside a Composio tool session; helpers to call any Composio tool, an LLM (invoke_llm), web search, file extraction and direct API proxy; pandas, numpy, matplotlib, PyTorch preinstalled, other packages auto-installed; auto-corrects common code mistakes; experimental files API; Python and TypeScript SDKs; separate experimental local sandbox |
| Platform feature | Platform feature | $0.582 | — | Workspace files persist across sessions and restarts; no idle or sleep rule published | Container | Feb 2026 | Agent platform pitched as 'Claude Code in the cloud as an API': each agent gets a persistent cloud computer withAgent platform pitched as 'Claude Code in the cloud as an API': each agent gets a persistent cloud computer with browser, terminal, files and optional GUI desktop; projects, tasks, memory, skills, subagents; triggers by email, cron schedules and webhooks; TypeScript and Python SDKs (npm computer-agents, pip computer-agents); integrations Google Drive, OneDrive, GitHub, Notion, Slack, Telegram; iOS and Mac apps; company in Germany |
Conductor Cloud (Conductor) Platform featureJoining Meta: 0.93.0 (2026-11-04) announced Cloud going read-only 2026-11-13 and shuttingJoining Meta: 0.93.0 (2026-11-04) announced Cloud going read-only 2026-11-13 and shutting down 2026-11-20; 0.94.2 (2026-11-13) says Conductor isn't shutting down and removed the Cloud shutdown banner | Platform feature | Cloud workspaces need a Pro, Teams (invite-only) or Enterprise planCloud workspaces need a Pro, Teams (invite-only) or Enterprise plan; the plan prices aren't on the public site (/pricing redirects to the homepage); no extra fee for cloud compute 'right now', with usage-based pricing planned | No separate cloud fee today | A cloud workspace sleeps after 20 minutes without agent or terminal activityA cloud workspace sleeps after 20 minutes without agent or terminal activity; files and chat history survive sleep, running processes don't; opening the workspace wakes it; every sandbox stops at a maximum lifetime of 23 h 50 min even if active | VM | Jun 2026 | Mac desktop app, iOS app and API that run Claude Code, Codex, Cursor and OpenCode with your own subscriptions or keysMac desktop app, iOS app and API that run Claude Code, Codex, Cursor and OpenCode with your own subscriptions or keys; cloud sandbox 8 vCPU, 16 GB RAM, 32 GB ephemeral NVMe, Amazon Linux 2023 with Node, Python and Chrome, us-east-1; repos and dependencies preinstalled through a configurable Cloud Computer; multiplayer shared workspaces; GitHub repos only; transfer cloud workspaces to local; no bring-your-own cloud yet (self-hosted offering in development); built with Vercel; SOC 2 Type II |
| Platform feature | Platform feature | Not priced separately: Cortex Agents bill orchestration in AI credits per million tokensNot priced separately: Cortex Agents bill orchestration in AI credits per million tokens (Service Consumption Table 6(d)); the docs say the code execution tool can add cost beyond orchestration and point to the Service Consumption Table (effective 2026-11-12), which lists no code-execution or sandbox rate | — | Sandbox scoped to one conversation thread and reused across requests on itSandbox scoped to one conversation thread and reused across requests on it; suspends after a period of inactivity (duration not stated); in-memory state isn't kept between separate code executions; files in the mounted workspace stage persist | — | Aug 2026 | Built-in Cortex Agents tool (code_execution) that runs Python 3.12 with numpy, pandas, scipy, pyarrowBuilt-in Cortex Agents tool (code_execution) that runs Python 3.12 with numpy, pandas, scipy, pyarrow, matplotlib and plotly preinstalled; more PyPI packages through the Artifact Repository; read/write access to the user's workspace stage; sandbox SQL limited to RESULT_SCAN on queries the agent already ran, with no table access; also runs Python scripts in agent skills; code_toolset_all gives the full Cortex Code toolset (bash, file edit, grep, web search, SQL); not available to agents called with owner's rights |
| Platform feature | Platform feature | No hosted service or published prices: self-host for free with Docker Compose or HelmNo hosted service or published prices: self-host for free with Docker Compose or Helm (OSS edition); Enterprise edition (SSO, fine-grained roles, audit logs, trace viewer, cost insights, multiple orgs) unlocked by a license key, price not published | N/a (self-hosted) | A sandbox keeps its files and installed packages across container restartsA sandbox keeps its files and installed packages across container restarts; a long-idle container shows Paused and resumes on next use (Docker pause, no checkpoint; the Docker Compose setup defaults pause_on_idle to false); Restart stops the container and keeps the storage; Off restarts on the next message; Delete removes the sandbox for good; each run has an execution timeout, and commands over 60 s become background tasks | Container | Jul 2026 | Self-hosted workspace platform for long-lived, team-owned agents: one workspace agent with role, memorySelf-hosted workspace platform for long-lived, team-owned agents: one workspace agent with role, memory, skills and approved MCP tools; reachable from web/H5 and IM (Slack, Discord, Teams, Feishu, DingTalk, WeCom per the lead); persistent per-workspace sandboxes; admin network policies (default allow or deny, host rules) and command rules (allow, deny, human confirm); sandbox secrets vault; agent runtime is CubeLoop (MIT); bring your own model providers; GitHub cubeplexai/cubeplex 576 stars; docs say OSS edition is Apache-2.0 while the repo LICENSE is the CubePlex Community License (self-host and commercial use allowed, CubePlex branding must stay); related new repo cubepod (persistent rootfs microVM platform, 4 stars) |
| Platform feature | Platform feature | Cloud computer prices aren't shown without signing in: Pro 2 vCPU / 8 GB RAMCloud computer prices aren't shown without signing in: Pro 2 vCPU / 8 GB RAM, Business 4 vCPU / 16 GB RAM; the desktop app for your own computers is free | — | — | — | — | Runs AI agents (Claude Code CLI, Pi) on your own Mac, Windows or Linux machines or Android phone through a free appRuns AI agents (Claude Code CLI, Pi) on your own Mac, Windows or Linux machines or Android phone through a free app, or on a Dashob cloud computer: a full Ubuntu XFCE desktop with Chrome, VS Code, LibreOffice, terminal and remote desktop that launches in 3-5 minutes; shared AI Drive; bring your own Anthropic, OpenAI, Gemini or OpenRouter keys (AES-256-GCM encrypted); use from a browser or phone; cloudbot-ai.com redirects to dashob.com |
| Platform feature | Platform feature | $0.1656 | $0 while paused | Compute bills while the environment runs, including while it waits for the next messageCompute bills while the environment runs, including while it waits for the next message; a paused environment is not billed; per-environment timeout (docs example 30m); sandbox torn down at session end, with its GitHub credentials revoked and source deleted | VM | Feb 2024 | Managed cloud sessions for Claude Code, Codex, Gemini CLI and OpenCode; agents defined in YAML, Python or TypeScriptManaged cloud sessions for Claude Code, Codex, Gemini CLI and OpenCode; agents defined in YAML, Python or TypeScript; triggers from GitHub pushes/PRs, Slack, Linear, Sentry, Jira and webhooks; API, CLI, Python and TypeScript SDKs, MCP server; reusable environments with cached build_base images; CPU/memory set per environment; per-session scoped GitHub credentials; budgets per session/agent/account; SOC 2 Type I (Type II in progress); full audit trail |
| Platform feature | Platform feature | Early access; no prices publishedPricing page | — | — | — | Nov 2026 | Emdash 2.0: coding agents for whole teams (Claude Code, Codex, CursorEmdash 2.0: coding agents for whole teams (Claude Code, Codex, Cursor, OpenCode or any ACP agent) on your own infrastructure, on Emdash cloud or locally; bring your own subscriptions, models and cloud commitments; start tasks from Slack, Linear, GitHub or a phone; shared environments, skills, MCP servers and secrets; SSO, roles, spend limits. Emdash Classic, the earlier open-source desktop app (macOS, Windows, Linux), runs agents in parallel git worktrees locally or on remote machines over SSH (Apache-2.0, 6,252 GitHub stars) and now gets only fixes; made by General Action |
| Platform feature | Platform feature | Cloud agents pricing not published; Umans separately sells open-model inference perCloud agents pricing not published; Umans separately sells open-model inference per token, e.g. DeepSeek V4.1 Flash $0.15/M input, $0.60/M output | — | Not stated for Cloud agents; the January 2026 envs page described snapshots and forks ofNot stated for Cloud agents; the January 2026 envs page described snapshots and forks of a running sandbox | — | Jan 2026 | Envs.umans.ai now redirects to app.umans.ai/offers/envsEnvs.umans.ai now redirects to app.umans.ai/offers/envs, which shows 'Cloud agents by Umans' and says 'Cloud agents is the next chapter of envs': connect a GitHub or GitLab repo and an agent plans, edits, runs your tests and builds in a real environment, driven from the browser. The January 2026 envs waitlist page (archived) described repo-ready sandboxes: environments built from devcontainer.json or a Dockerfile and cached, warm starts under 2 s, snapshot and fork, streaming exec, MCP server, egress allow list (default blocks all but package registries), secrets injection, command audit, VPC data plane in AWS/GCP for enterprise. Umans AI ('Made in Europe') now mainly sells open-model LLM inference |
| Platform feature | Platform feature | Prepaid credits from $5, no subscription: $0.005 run fee + $0.0002 per second of run timePrepaid credits from $5, no subscription: $0.005 run fee + $0.0002 per second of run time ($0.72/hour) + model calls at provider price +15% + web $0.001/request and $0.015/search; zero data retention +10%; budget per run up to $100 (default $0.50); VM size not published. Self-hosted in your AWS: Team $490/month (1 installation, 10,000 runs/month), Business $5,000/month | N/a (no pause; the VM is destroyed after each run) | No pause: each run gets a fresh microVM that is destroyed when the task endsNo pause: each run gets a fresh microVM that is destroyed when the task ends; timeout 5 min by default, at most 1 hour; run records kept 30 days | VM | Oct 2026 | Runs a whole agent task (prompt, input, files, allowed HTTP APIs, MCP servers and websites, modelRuns a whole agent task (prompt, input, files, allowed HTTP APIs, MCP servers and websites, model, budget) in Ephemeral's own worker; you do not bring your own agent loop; API, CLI, MCP server and dashboard; saved workflows and schedules; signed callbacks; files mounted read-only, 1 MiB each and 4 MiB per run; shell can be turned off; sites in 7 languages; produced by erni.works |
| Platform feature | Platform feature | Free, open source (MIT); no paid edition; runs on your own Docker EnginePricing page | N/a (self-hosted) | Each create, exec and start refreshes a 20-minute lease on the containerEach create, exec and start refreshes a 20-minute lease on the container; stop pauses the container and keeps its filesystem; remove deletes the container, its network and every file not downloaded | Container | Apr 2026 | Capability of the Everruns open-source Rust agent framework and durable runtimeCapability of the Everruns open-source Rust agent framework and durable runtime; one self-hosted Linux container per agent session via the Docker Engine API; shell exec, file I/O, file transfer to session storage; Daytona and E2B are separate hosted-sandbox capabilities; self-host on a machine, Docker Compose or Kubernetes; non-commercial project with no company behind it; MIT; 67 stars |
| Platform feature | Platform feature | Private preview through Factory's account team; you pay for your own infrastructurePricing page | N/a (self-hosted on your infrastructure) | Tracks each sandbox's activity and suspends or terminates it according to policy and whatTracks each sandbox's activity and suspends or terminates it according to policy and what the provider supports (optional Suspend, Snapshot and Fork conformance profiles per driver) | VM | Oct 2026 | Self-hosted orchestrator for Factory Droid agent sandboxes in your own cloud accounts or data centersSelf-hosted orchestrator for Factory Droid agent sandboxes in your own cloud accounts or data centers; backend for Factory Environments, Automations, Delegations and Remote Control; connects outward to Factory over TLS WebSocket (no inbound port), with a key that can sit in a cloud KMS; each sandbox makes its own outbound connection so private repos stay private; default for Factory Private; pluggable drivers with conformance tests; works only with Factory's control plane |
| Platform feature | Platform feature | No extra charge for code execution; billed as the model's input and output tokensNo extra charge for code execution; billed as the model's input and output tokens (generated code and its results count as tokens) | N/a (no persistent sandbox) | Runs only inside a model call: each code run has a maximum runtime of 30 secondsRuns only inside a model call: each code run has a maximum runtime of 30 seconds, and the model may retry up to 5 times on error; no persistent sandbox described | — | Jun 2024 | Built-in tool that lets Gemini models write and run Python onlyBuilt-in tool that lets Gemini models write and run Python only; fixed library set of about 37 packages (numpy, pandas, scipy, scikit-learn, tensorflow, matplotlib, seaborn, opencv-python, geopandas, python-docx and others), and you can't install your own; file input (CSV, text) via inline data or the Files API, graph output as inline images; combines with Google Search grounding and function calling on Gemini 3 and 4 models |
| Platform feature | Platform feature | Compute $0.000024 per vCPU-second (overview's unit) or per compute second = sessionCompute $0.000024 per vCPU-second (overview's unit) or per compute second = session runtime (billing page; its example bills 600 s with 4 GiB as $0.0144 + $0.0072 with no vCPU count); memory $0.000003/GiB-second allocated; snapshot storage $0.005/GiB-month for stopped sessions; usage draws first from the plan's included Actions compute entitlement (now a shared dollar-based pool); billing begins November 2026; local sandboxing included in the Copilot seat. 2 vCPU/4 GiB would be $0.2160/hr if compute is per vCPU-second, $0.1296/hr if per session-second; sandbox vCPU/RAM size not published | $0 compute and memory while stopped; snapshot storage $0.005/GiB-month | When a session stops or the sandbox shuts down, it saves a snapshot of its stateWhen a session stops or the sandbox shuts down, it saves a snapshot of its state; resuming restores files and in-progress work from any surface; compute and memory not metered while stopped, snapshot storage metered until deleted; no one-hour limit on total session time; idle timeout not stated | Process | Jun 2026 | Default environment for Copilot cloud sessions started from github.com, GitHub Mobile, Slack, Microsoft TeamsDefault environment for Copilot cloud sessions started from github.com, GitHub Mobile, Slack, Microsoft Teams, the Copilot app, VS Code and the Copilot SDK; Copilot CLI uses it with copilot --cloud; network follows Copilot cloud agent firewall allow-list; org/enterprise owner must enable the Cloud Sandbox access policy; Copilot cloud agent with Actions-based setup still runs in GitHub Actions |
| Platform feature | Platform feature | Gobii.ai/pricing unreachable because gobii.ai has no DNSGobii.ai/pricing unreachable because gobii.ai has no DNS; self-hosted platform free, open source (MIT) | — | A per-agent sandbox pod stops after 1 hour idle (idle-only TTL)A per-agent sandbox pod stops after 1 hour idle (idle-only TTL); at idle stop its 1 GB workspace volume is snapshotted to disk and restored on the next call; no memory snapshot (design spec v1) | gVisor | Aug 2025 | 'AI employee' platform: always-on agents based on browser-use that wake on schedules, email, SMS and webhooks'AI employee' platform: always-on agents based on browser-use that wake on schedules, email, SMS and webhooks, each with its own email address and phone number; headed browsers with persistent profiles and proxy routing; agent-to-agent messaging; SQLite operational memory; encrypted secrets; REST API and remote MCP for Claude Code, Cursor and VS Code; self-host with Docker Compose (12 GB RAM); the README points to a hosted Gobii Cloud at gobii.ai; 500 GitHub stars |
| Platform feature | Platform feature | Not stated; the Grafana Cloud pricing page doesn't mention Grafana Assistant or ManagedNot stated; the Grafana Cloud pricing page doesn't mention Grafana Assistant or Managed Sandboxes | — | — | — | Oct 2026 | Grafana Assistant hands a diagnosis built from metrics, logs, traces and profiles to a coding agent that edits codeGrafana Assistant hands a diagnosis built from metrics, logs, traces and profiles to a coding agent that edits code, runs tests and builds, and commits to a branch; diff shown in Grafana with Create PR; per-repository publish policy (ask before publishing, or auto-publish draft or ready PRs for verified private repos); GitHub App; per-repository environments with network restrictions and secrets; extra tools via .grafana/Dockerfile; sandboxes run on AWS in the EU or US by stack region; BYOC not supported in the preview; can hand tasks to Cursor Cloud or Claude Managed Agents on your own account instead; @grafana mentions in GitHub issues and PRs; scheduled automations |
| Platform feature | Platform feature | Free $0 (250 credits a month, basic sandbox)Free $0 (250 credits a month, basic sandbox); Pro $20/mo or $17 billed yearly (2,000 credits, sandbox 2 cores, 4 GB RAM, 50 GB); Plus $50/mo or $42 (5,000 credits, same sandbox); Max $200/mo or $167 (22,000 credits, sandbox 4 cores, 8 GB, 200 GB); credits cover AI usage; no hourly sandbox rate | — | Projects are persistent workspaces for sessions and files | — | Mar 2026 | Consumer 'agent-native computer': frontier models, a skill storeConsumer 'agent-native computer': frontier models, a skill store, scheduled automations and email (Capymail) on a cloud sandbox; Connect My Mac for local files; iOS app |
| Platform feature | Platform feature | Instance billed daily in arrears from Nous credit: running = compute per second of uptimeInstance billed daily in arrears from Nous credit: running = compute per second of uptime + storage; stopped = storage only; per-day size prices shown only inside Nous Portal (sign-in); $2 minimum credit or a subscription to deploy; every paid plan includes one Cloud agent: Plus $20/mo (up to 4 GB RAM, 25 GB disk), Super $100/mo (4 GB, 100 GB), Ultra $200/mo (8 GB, 100 GB); Free plan pay-as-you-go; inference and tools billed separately | Storage only when stopped (rate shown only in Nous Portal) | Site says it 'scales to zero when idle. You only pay while it works'Site says it 'scales to zero when idle. You only pay while it works'; compute is metered to the second of uptime; stopping the agent drops it to a storage-only rate and keeps its persistent workspace, memory and configuration | Container | Jul 2026 | One-click deploy of Nous Research's open-source Hermes Agent (MIT) on a dedicated instance with its own workspace andOne-click deploy of Nous Research's open-source Hermes Agent (MIT) on a dedicated instance with its own workspace and dashboard; reach it from Telegram, Discord, Slack, email and CLI; natural-language schedules; parallel subagents; manage instances (list, start, stop) from a local Hermes Agent over the Nous Portal MCP server; sizes can only grow (disks never shrink); locations Amsterdam, Stockholm, Mumbai, Paris, Dallas, Frankfurt, Los Angeles, London, Tokyo, Chicago, Singapore, San Jose, Sydney, Toronto |
| Platform feature | Platform feature | Seats: Free $0 (5 concurrent cloud sessions, sandboxes up to 2 vCPU / 8 GiB, 5 projects)Seats: Free $0 (5 concurrent cloud sessions, sandboxes up to 2 vCPU / 8 GiB, 5 projects); Team $60/seat/month or $50 yearly (150 concurrent sandboxes per seat, up to 16 vCPU / 64 GiB); Enterprise custom. Plans include model and sandbox usage; beyond that, prepaid credits (1 credit = $1). Sandbox rates per hour by size: X Small 1 vCPU / 4 GiB $0.08; Small 2 vCPU / 8 GiB $0.16; Medium 4 vCPU / 16 GiB $0.30; Large 8 vCPU / 32 GiB $0.75; X Large 16 vCPU / 64 GiB $1.50; no 2 vCPU / 4 GiB size | $0 between runs (idle time and sandboxes left running after a preview or$0 between runs (idle time and sandboxes left running after a preview or terminal are free); archived workspace storage ? | The workspace stays with the thread across follow-upsThe workspace stays with the thread across follow-ups; an idle workspace is archived and restored when the thread continues; if a runtime fails, Hoplite restores its latest snapshot; Modal sandboxes are billed only from run start until 5 minutes after the run ends | VM | Aug 2026 | Cloud coding-agent platform (Claude Code or Codex, with your own Claude or ChatGPT subscription or Hoplite's models)Cloud coding-agent platform (Claude Code or Codex, with your own Claude or ChatGPT subscription or Hoplite's models); each thread's sandbox gets the repo, encrypted env vars and a setup script from .hoplite/settings.json; private Docker daemon for Docker Compose, Supabase or Testcontainers repos (Compose VM 4 vCPU / 8 GiB); browser checks and live previews; orchestrator with parallel subagents; PR auto-fix; Slack, Linear and Sentry triggers; macOS and iOS apps, CLI, MCP server and Platform API; short-lived GitHub tokens |
| Platform feature | Platform feature | Hosted-agent computer pricing is not published; the one size is 1 vCPU / 4 GiBHosted-agent computer pricing is not published; the one size is 1 vCPU / 4 GiB; identity plans: Free $0 (3 identities), Developer $25/mo or $300/yr (10), Startup $167/mo or $2,000/yr (100), Enterprise custom; SMS/MMS overage $0.03/message, calls $0.03/minute | — | The computer stops after 1 or 5 minutes idle (default 1 minuteThe computer stops after 1 or 5 minutes idle (default 1 minute; can't be turned off), keeping files, conversations and schedules; selected incoming messages (email, SMS, iMessage, WhatsApp, Slack) and schedules start it again, even after a manual stop; removing hosting deletes the computer and its files | — | Nov 2026 | Hosted agents launched 2026-11-09: choose the Hermes or Codex harness and the agent gets its own cloud computer tied toHosted agents launched 2026-11-09: choose the Hermes or Codex harness and the agent gets its own cloud computer tied to its Inkbox identity (email inbox, phone number, iMessage, WhatsApp, Slack, agent-to-agent); terminal access, with packages and files kept on its disk; schedules once, on an interval or cron; chat API with Ask or YOLO permission modes; default model gpt-6.1-terra or your own key; sandboxes for agents you run elsewhere announced as coming soon; Python, TypeScript and Rust SDKs, CLI, MCP |
KiloClaw (Kilo)Shut down: end of life, closed to new users and support ending Platform feature | Platform feature | No public price now: kilo.ai/pricing no longer lists KiloClawNo public price now: kilo.ai/pricing no longer lists KiloClaw. Kilo's billing code (Kilo-Org/cloud kiloclaw-pricing-catalog.ts): price version 2026-05-10 Standard $55 recurring and $306 for a six-month commit, 1-day trial; version 2026-03-19 Standard $9 ($4 intro), $48 six-month commit, 7-day trial; models billed from Kilo Gateway credits or your own key | — | Instances ran until stopped; Start Machine boots a stopped instance in up to 60 sInstances ran until stopped; Start Machine boots a stopped instance in up to 60 s; the persistent volume stays in the region where the instance was created; Kilo's billing code can suspend, stop and later destroy an instance after its subscription lapses | VM | Feb 2026 | Hosted OpenClaw on one dedicated machine per user with 2 shared vCPUsHosted OpenClaw on one dedicated machine per user with 2 shared vCPUs, 3 GB RAM and a 10 GB persistent SSD (archived docs; the code's default guest is performance-1x, 1 CPU, 3 GB); OpenClaw bound to loopback behind a Kilo controller; per-user gateway token; secrets encrypted at rest; Telegram, Discord and Slack; GitHub, Google, Linear, Composio, 1Password and Brave Search tools; webhook and scheduled triggers; version pinning |
| Platform feature | Platform feature | CNY only: sandbox ¥0.072 per hour of run time per session (no sandbox size stated)CNY only: sandbox ¥0.072 per hour of run time per session (no sandbox size stated); plus model tokens at model rates and per-call plugin fees; self-hosted environments run the sandbox tools on your own hosts with no sandbox charge | ¥0 sandbox charge while the session is idle | Sessions are idle, running or terminated; sandbox time stops accruing once all threadsSessions are idle, running or terminated; sandbox time stops accruing once all threads finish and the session goes idle; files written to /mnt/agents/ sync to a persistent workspace and are remounted if the sandbox instance is rebuilt; archive makes a session read-only, then it can be deleted | — | Sep 2026 | Managed agent harness (Kimi Durable Harness) with built-in shell, file and todo tools running in the sandboxManaged agent harness (Kimi Durable Harness) with built-in shell, file and todo tools running in the sandbox; Python 3.12, Node.js 24, .NET 8, Go, Rust, LibreOffice, pandoc, headless Chromium, ffmpeg and Tesseract preinstalled; environments set packages, network policy (proxy, host allowlists) and init scripts; skills, MCP, plugins, credential vaults, memory stores with Dreams, multi-agent orchestration, cron triggers; kha CLI; self-hosted environments for enterprise-verified users |
| Platform feature | Platform feature | $0.108 | $0 compute while stopped; storage $0.0526/GB-month | The computer sleeps about 15 to 20 minutes after the last turn and keeps /workspaceThe computer sleeps about 15 to 20 minutes after the last turn and keeps /workspace, the session's branch and its messages; a new message or preview request wakes it; the home directory, /tmp and installed packages can be gone after a wake | — | Oct 2024 | Open-source 'CompanyOS' agent harness: agents, skills and memory live in one git repoOpen-source 'CompanyOS' agent harness: agents, skills and memory live in one git repo; each session gets a cloud computer with terminal, browser, a desktop you can watch, Python and Node; sandbox templates set software and size; connect your own computer; self-host with Docker Compose; web, desktop and mobile apps plus Slack and Teams; GitHub kortix-ai/CompanyOS, 20,724 stars |
| Platform feature | Platform feature | Pricing is behind sign-in | — | — | — | Jun 2023 | AI-native workspace for planning and building software with coding agents: starts Claude Code sessions in sandboxedAI-native workspace for planning and building software with coding agents: starts Claude Code sessions in sandboxed worktrees per issue, drafts issues with Claude Code, keeps codebase and product context that learns from issues, chats and reviews, and offers a code-review UI for AI-generated changes; marketing copy renders client-side and /pricing and /docs redirect to sign-in; the domain launched in 2023 as an AI coding assistant |
| Platform feature | Platform feature | $0.0411 | Same as running (flat monthly fee) | Always on: a persistent VM that keeps filesAlways on: a persistent VM that keeps files, tools and running processes between Manus sessions; no suspend; an upgrade restarts the VM; if the subscription lapses the VM is shut down and its files, tools and databases are deleted | VM | Apr 2026 | Persistent Ubuntu server rented inside the Manus agent app so Manus can run websites, bots, scripts and databases 24/7Persistent Ubuntu server rented inside the Manus agent app so Manus can run websites, bots, scripts and databases 24/7; command line only, no GUI; Manus can deploy one automatically when a task needs persistence; separate from the per-task temporary sandbox and from "My Computer" (Manus desktop app on your own machine) |
| Platform feature | Platform feature | $1.50 | $0 compute once the idle sandbox is destroyed$0 compute once the idle sandbox is destroyed; checkpoint storage price not stated | Platform destroys a sandbox after idleTimeoutMinutes with no activity (docs examples usePlatform destroys a sandbox after idleTimeoutMinutes with no activity (docs examples use 30); stop() releases the lease while the VM keeps running under the environment's idle policy; with a stable id a new sandbox boots from that id's last filesystem checkpoint, and snapshot() saves one at once | VM | Jul 2026 | Sandboxes inside Mastra Platform, the hosting for agents built with the Mastra TypeScript frameworkSandboxes inside Mastra Platform, the hosting for agents built with the Mastra TypeScript framework; each environment gets a file bucket plus on-demand sandboxes; template with Python 3, Node 22, TypeScript; Template() builder sets cpuCount, memoryMB and Docker (E2B); private networking and public ports; also LocalSandbox (not isolated by default) and RailwaySandbox; Mastra Factory uses sandboxes for coding-agent sessions; framework 29,526 GitHub stars, core Apache-2.0 with ee/ directories under a separate license |
| Platform feature | Platform feature | No public rates. Billed by elapsed time as hours x (per_hour + reserved_memory_GiB xNo public rates. Billed by elapsed time as hours x (per_hour + reserved_memory_GiB x per_gib_hour) against a per-deployment rate card (SKUs standard, browser, eval-python) that only cluster admins can read or set | Rate card per deployment; stopped sessions keep workspace 7 days | ttl_seconds is an idle timeout (default 600 s), renewed by every commandttl_seconds is an idle timeout (default 600 s), renewed by every command; after it the container is reclaimed and the session is stopped; /workspace files persist 7 days; next command re-provisions and rehydrates in several seconds | Container | — | Part of MeetKai's MKA1 "sovereign AI" platform; standalone sandbox API plus shell and code_interpreter tools in itsPart of MeetKai's MKA1 "sovereign AI" platform; standalone sandbox API plus shell and code_interpreter tools in its Responses API; persistent /workspace; headless-Chrome browser sessions with CDP; only resource_class medium accepted (example reserves 256 MiB); queue_if_full option; TypeScript, Python and C# SDKs |
| Platform feature | Platform feature | Free to start, no credit card; models run on your own provider keyFree to start, no credit card; models run on your own provider key; no paid plans published | Sandboxes are reclaimed between turns; no price published | The Session outlives its sandbox: every successful turn ends in a checkpointThe Session outlives its sandbox: every successful turn ends in a checkpoint, the sandbox is reclaimed, and the next turn (minutes or days later) restores the working directory and conversation, with a longer cold start; a turn runs up to 2 hours; sandboxes are capped per account and a job past the cap fails at once | Container | Jun 2026 | API-first managed agent runtime for application backends: runs OpenAI Codex, Claude Agent SDKAPI-first managed agent runtime for application backends: runs OpenAI Codex, Claude Agent SDK, OpenCode or Pi behind one REST API with SSE event streams, files and artifacts; Agent presets with instructions, Skills, remote MCP servers and pinned npm/PyPI packages; full or allowlisted network; CLI and agent skill; separate product mosoo computer (launched 2026-09-21) is a persistent cloud computer that runs Codex, Hermes or OpenClaw, reachable from Telegram, Slack, Discord or Feishu/Lark, and sleeps by stopping its container while keeping files and sessions; open source (Apache-2.0), 166 GitHub stars |
| Platform feature | Platform feature | The app needs a sign-in; no public prices | — | Sleeps by stopping its container; keeps workspace files, runtime state and sessions | Container | Sep 2026 | A persistent cloud computer for your agent (Codex, Hermes or OpenClaw) that you reach from chat appsA persistent cloud computer for your agent (Codex, Hermes or OpenClaw) that you reach from chat apps; by LangGenius, the company behind Dify |
| Platform feature | Platform feature | Nebula Workspace $25/seat/mo (7-day trial with card), includes one Small cloud deviceNebula Workspace $25/seat/mo (7-day trial with card), includes one Small cloud device; more cloud-device time draws Nebula Credits while running (rate shown in the workspace, not on the public pages); sizes Small 2 CPU/8 GB/40 GB, Medium 4/16/250, Large 8/32/500; first workspace gets $3 of credits and 1,000 messages | Credits are drawn while a device runs; stopped cost not stated | Cloud devices are Running or Stopped; a new device starts Stopped and boots when an agentCloud devices are Running or Stopped; a new device starts Stopped and boots when an agent first uses it; without a subscription a device unused for 60 days is deleted with its files; with a very low credit balance, a device not used by a live channel or agent can be deleted after 14 idle days | — | — | Cloud devices are the computers agents work on inside Nebula, a team chat workspace for people and agentsCloud devices are the computers agents work on inside Nebula, a team chat workspace for people and agents; channels pick a default device; run Claude Code, Codex, Gemini CLI or Cursor CLI on a cloud device or your own computer; terminal, files and services tabs; SSH keys log in as root; device sharing per person or workspace; apps on macOS, Windows, Linux, iPhone, Android |
| Platform feature | Platform feature | No prices on the site or in the README; users bring their own model-provider keysNo prices on the site or in the README; users bring their own model-provider keys; a self-hosted copy needs your own E2B key | — | Sessions are durable server-side objects that survive closed tabs and redeploysSessions are durable server-side objects that survive closed tabs and redeploys; the live runtime is released when idle and rebuilt on demand; E2B sandboxes are idle-paused with state preserved | VM | Jan 2026 | Multi-user browser agent platform, installable as a PWA, built on the pi-mono agent engineMulti-user browser agent platform, installable as a PWA, built on the pi-mono agent engine; per-session execution target including machines with no public IP; subagents, persisted agent trees, task queues, teams and declarative workflows; three approval policies; MCP servers as tools; share links that redact secrets; audit log; a security-assessment profile with recon tooling in its E2B template; Node/Express with SQLite; GitHub n-WN/SubgroupX 375 stars, no license file, last push 2026-08-04 |
| Platform feature | Platform feature | Credits: Free $0 with 1,000 credits valid 30 days (2 concurrent runs)Credits: Free $0 with 1,000 credits valid 30 days (2 concurrent runs); Pro from $20/month (member packages $20-$200); Team from $180/month (plan $160 + member packages); Enterprise custom; top-up $1 per 1,000 credits; credits not tied to vCPU-hours | — | Each agent run gets an isolated environment that is destroyed when the run finishes | VM | Nov 2025 | vm0.ai redirects to okou.ai ('VM0 is now Okou. Same team, same product, new name.')vm0.ai redirects to okou.ai ('VM0 is now Okou. Same team, same product, new name.'); now an AI teammate for team workflows (Slack, Gmail, GitHub, Notion, Linear and other connectors) that produces reports, decks and websites; connect a Claude or Codex subscription or bring your own keys; credentials brokered outside the agent; source on GitHub okou-ai/okou (1,163 stars, license NOASSERTION) |
| Platform feature | Platform feature | $0.1656 | $0 runtime while asleep; retention $0.20016 per GiB of memory per 30 days$0 runtime while asleep; retention $0.20016 per GiB of memory per 30 days (~$0.80/month for 4 GiB) | Idle Omnara-managed machines auto-sleep, keeping data and pausing runtime charges (BlaxelIdle Omnara-managed machines auto-sleep, keeping data and pausing runtime charges (Blaxel machines resume in about 25 ms with memory, files and processes); pool machines are deleted when released or after a configured idle period | VM | Oct 2026 | Open-source managed agent platform, a self-hostable alternative to Claude Managed Agents and OpenAI's Agents APIOpen-source managed agent platform, a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API: agents defined in YAML (model, tools, approvals, machine sources); the agent loop runs in Omnara's control plane and each agent is an event log that attaches and releases machines; the omnarad daemon connects outbound, no inbound ports; Claude Code and Codex can use Omnara as an MCP server; Slack integration; Apache 2.0, 3,396 GitHub stars; SOC 2 Type II observation period in progress |
| Platform feature | Platform feature | $0.36 | Codex cloud: included in plan; Agents API: ? (billed per session/minute whileCodex cloud: included in plan; Agents API: ? (billed per session/minute while the container exists) | Codex cloud: a task's saved VM state is recoverable for up to 7 days after the last turnCodex cloud: a task's saved VM state is recoverable for up to 7 days after the last turn or resume. Agents API: files persist across turns while the sandbox exists; deleted after 1 hour without activity or keep-alives (not configurable); files in /workspace/outputs kept as artifacts after expiry | VM | May 2025 | Codex cloud: reusable environments with repos, dependencies, install script and start skill, internet allow-listsCodex cloud: reusable environments with repos, dependencies, install script and start skill, internet allow-lists, Enterprise larger/custom VMs. Agents API: sandbox sizes small 1 vCPU/1 GB, medium (default) 2 vCPU/4 GB, large 4 vCPU/16 GB; Python, Node.js and command-line tools preinstalled; environment templates; works only with OpenAI's Agents API (beta SDK) |
| Platform feature | Platform feature | Billed in work minutes (time the agent is working, per secondBilled in work minutes (time the agent is working, per second; waiting is free): 3-day trial 600 minutes incl. $10 model tokens; Pro $20/mo for 2,000 minutes; Max $200/mo for 25,000 minutes; extra minutes $0.01 from prepaid credits; Enterprise custom; bring your own Claude/ChatGPT plan or API key, or OpenComputer tokens at API price; self-hosted free (Apache-2.0). No per-machine rate is published | $0 (waiting time is not billed in work minutes) | A thread's runner exits after 2 h idle and resumes on the next message (suspendedA thread's runner exits after 2 h idle and resumes on the next message (suspended sessions still answer reads); idle session directories removed after 14 days if nothing is uncommitted or unpushed, build/dependency directories after 24 h | VM | Dec 2025 | Pivoted 2026-11-13 from general-purpose sandboxes to @ComputerPivoted 2026-11-13 from general-purpose sandboxes to @Computer, an open-source Devin alternative that works from Slack threads; runs Claude Code, Codex or OpenCode sessions; self-host on AWS, Azure or Google Cloud (one VM per org plus Cloudflare Workers gateway); Computer API over WebSocket/HTTP and opc CLI; preview URLs; Serverless Agents in TypeScript run in OpenComputer cloud; GitHub diggerhq/opencomputer Apache-2.0, 635 stars; made by Digger |
| Platform feature | Platform feature | Free, open source (MIT)Pricing page | N/a (self-hosted) | Each Worker is persistent and keeps its sandbox computer's state between conversationsEach Worker is persistent and keeps its sandbox computer's state between conversations; no idle policy stated | Container | Jul 2026 | Self-hosted platform for OpenWorker, a conversational desktop agentSelf-hosted platform for OpenWorker, a conversational desktop agent; each Worker gets its own Docker sandbox with XFCE, Chrome, VS Code, Obsidian; tiers chat, browser, desktop, playbook; MCP, CLI, REST, web and Discord/Telegram gateway; local Docker or remote SSH sandboxes; one OpenRouter key; 0 stars |
| Platform feature | Platform feature | $0.35 per RCU-hour; 1 RCU = 1 CPU core + 4 GB memory; scales to 8 cores / 32 GB (8 RCUs)$0.35 per RCU-hour; 1 RCU = 1 CPU core + 4 GB memory; scales to 8 cores / 32 GB (8 RCUs); Cloud Build included; API key by request form. No 2 vCPU / 4 GiB unit: nearest are 1 RCU (1 core/4 GB) $0.35/h and 2 RCUs (2 cores/8 GB) $0.70/h | — | — | Container | Mar 2024 | Hosted OpenHands sandbox runtime "designed specifically for agent evaluation purposes"Hosted OpenHands sandbox runtime "designed specifically for agent evaluation purposes"; bash, web browsing and file editing for the OpenHands agent; selected in OpenHands with RUNTIME=remote plus API URL and key; OpenHands is open source (MIT, 93k stars) |
| Platform feature | Platform feature | $0.1656 | No pause state; sandboxes are torn down at session end | A sandbox is usually created when an environment session starts and torn down when itA sandbox is usually created when an environment session starts and torn down when it ends (or only for one tool call); billed for the time spent in the sandbox | Container | — | Platform to train and evaluate agents on 430+ RL environments through one API, built on the Open Reward StandardPlatform to train and evaluate agents on 430+ RL environments through one API, built on the Open Reward Standard; OpenReward Sandboxes are tied to an environment workspace and the docs say they are "not a general sandbox compute solution"; sizes from 0.1 vCPU/0.1 GiB to 4 vCPU/16 GiB plus an nvidia-l4 preset; custom container images; optional network blocking; bucket mounts; sidecars; environments can use E2B, Daytona or Modal instead; built by GR Inc; "from the team behind Papers with Code, Llama 2 and Llama 3" |
| Platform feature | Platform feature | Sandbox time $0.0001/second ($0.36/hr) with engine 'openrouter'Sandbox time $0.0001/second ($0.36/hr) with engine 'openrouter', from the request's first sandbox command until the response completes; a request that starts a new or sleeping container is billed at least 30 s; idle time between requests not billed; Files API usage included; model tokens billed separately; no CPU/RAM shape published | $0 (idle and sleeping containers not billed; saved files kept 30 days) | Container sleeps after 5 min idle (not configurable)Container sleeps after 5 min idle (not configurable); sleep keeps files in the home directory, and the next request starts a new sandbox and reloads them, but processes, env vars and installed system state are lost; saved container files kept 30 days after last use; container_auto containers are ephemeral per request or session, container_reference ids persist across requests; per-command timeout default 2 min, max 5 min | Container | Sep 2026 | Server tool that gives any tool-calling model a hosted shell on the Responses API (OpenAI-compatible shell) andServer tool that gives any tool-calling model a hosted shell on the Responses API (OpenAI-compatible shell) and Anthropic Messages API (openrouter:bash); Files API to attach inputs and download outputs (10 GiB workspace storage cap per third-party note, not checked); no outbound internet by default, domain allowlist up to 50 hosts, ports 80/443 only; up to 100 commands per call; global endpoint only (not eu./us. endpoints) |
| Platform feature | Platform feature | On-premise agent sandbox is an Enterprise Plan feature (contact sales)On-premise agent sandbox is an Enterprise Plan feature (contact sales); the code interpreter server tool has no separate charge in the current release | — | Sandboxes are claimed from a SandboxWarmPool of pre-provisioned pods and are short-livedSandboxes are claimed from a SandboxWarmPool of pre-provisioned pods and are short-lived; the orq:shell and orq:code_interpreter server tools create one sandbox per model response and discard it when the response ends (no cross-request containers); per-command timeout 60 s default, 1 to 300 s | Container | — | Runs customer-written Python (custom evaluators, agent code tools) and the orq:shell and orq:code_interpreter serverRuns customer-written Python (custom evaluators, agent code tools) and the orq:shell and orq:code_interpreter server tools inside the Orq.ai LLM platform and AI gateway; optional in the Helm chart, installed by the cluster administrator; sandbox-router routes executions to pods; runtime image ghcr.io/orq-ai/agent-sandbox-python-runtime (Python 3.13 on Debian, requests, pandas, pydantic) with a documented runtime contract (HTTP :8888, /execute, file staging under /workspace); air-gapped mirroring supported |
| Platform feature | Platform feature | $0.03 per sandbox session (one container; a session covers up to 20 minutes of active use$0.03 per sandbox session (one container; a session covers up to 20 minutes of active use as a billing window, not a runtime cap); SDK search queries made inside the sandbox $0.0025 each; other tools called from inside at their per-call rates (e.g. people_search $0.005); model tokens billed separately at Agent API rates; no CPU/RAM size published | — | One container per sandbox session; executions within one Agent API response share theOne container per sandbox session; executions within one Agent API response share the container (files and installed packages carry over between steps); each execution has a runtime cap (value not published, status timed_out); long jobs run with background: true and polling; persistence across responses not stated (the launch post described stateful sessions on a FUSE-mounted filesystem that pause and resume hours later, but the current docs don't) | Container | Mar 2026 | Server-side tool of Perplexity's Agent API: add {"type": "sandbox"} and the model writes and runs Python or bash itselfServer-side tool of Perplexity's Agent API: add {"type": "sandbox"} and the model writes and runs Python or bash itself (you don't supply the code); the code that ran comes back in sandbox_results with stdout, stderr and exit code; container has internet access and can install packages; preinstalled Perplexity SDK calls Web Search, Fetch URL Content and People Search from inside; GitHub connector (preview) lets git and gh use your connected credentials; files delivered with share_file and downloaded from the response files endpoint; output truncated around 1 MiB per stream; Python and TypeScript SDKs; launch post says it is the same execution environment Perplexity uses for Computer, Finance Agent and Deep Research |
| Platform feature | Platform feature | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (AGPL-3.0); hosted service at pipod.dev 'coming soon', not available yet | — | A pod stops on its own after 15 idle minutesA pod stops on its own after 15 idle minutes; pipod stop or pipod archive stop it at once; archives can be rehomed between hosts | Container | Oct 2026 | Runs pi coding-agent sessions in remote pods on a server you run; CLI (@pipod/cli) plus iOS and Android appsRuns pi coding-agent sessions in remote pods on a server you run; CLI (@pipod/cli) plus iOS and Android apps; control plane with REST API, session gateway, Postgres and Zitadel OIDC; per-pod ceiling default 8 vCPU / 24 GiB / 20 GiB; Docker Compose self-host on a Linux host with 8 GB RAM; AGPL-3.0; 167 stars; Billzo, LLC |
| Platform feature | Platform feature | ? on a first-party page (proliferate.com did not resolve on 2026-11-14)? on a first-party page (proliferate.com did not resolve on 2026-11-14). The repo's billing code sets Pro at $20 per seat a month with $15 of compute and $5 of LLM allocation per seat, 2 active environments per seat and 1 free trial hour, and prices managed compute per sandbox-hour as E2B's list price times a configured margin multiplier (no fixed rate in the code) | — | Cloud workspaces: E2B sandboxes created with a 2,700 s (45 min) timeoutCloud workspaces: E2B sandboxes created with a 2,700 s (45 min) timeout, which the server can pause and resume (per the repo's code) | — | Apr 2026 | Open-source desktop AI IDE (AGPL-3.0, 520 stars) that runs Claude Code, Codex, OpenCode, Cursor and Grok in parallelOpen-source desktop AI IDE (AGPL-3.0, 520 stars) that runs Claude Code, Codex, OpenCode, Cursor and Grok in parallel, each task in its own git worktree, locally or in the cloud; subagents; MCPs, skills, Computer Use and Browser Use; recurring and event-driven workflows; self-hostable control plane (Docker Compose, AWS CloudFormation, GCP, Azure, Kubernetes, air-gapped); Rust AnyHarness runtime |
| Platform feature | Platform feature | Prepaid Qoder Credits: model calls by token usePrepaid Qoder Credits: model calls by token use, plus sandbox runtime metered per second only while a Session is running (paused, idle and terminated not billed; charges began 2026-08-10); Qoder says sandbox runtime is typically 5-10% of a Session's credits; no credit rate per sandbox-hour published; storage free for a limited time; credits bought as Qoder Teams shared add-on packs or Service Account Dedicated Credits packs through Alibaba Cloud marketplaces | $0 (paused and idle Sessions not billed; storage free for a limited time) | Session states idle, running, rescheduling, terminatedSession states idle, running, rescheduling, terminated; the sandbox can pause, and a resumed sandbox keeps its files but not its processes (startup_script reruns on every resume); container storage may be reclaimed after 24 consecutive hours of inactivity, after which the container is reinitialized and unsaved files may be lost; archive or terminate ends a Session permanently; beta notes say sessions can run up to 26 hours | Container | May 2026 | Managed agent runtime from Qoder (Alibaba's AI coding product): define an Agent (model, prompt, tools)Managed agent runtime from Qoder (Alibaba's AI coding product): define an Agent (model, prompt, tools), an Environment (managed cloud or self_hosted) and a Session over REST, events over SSE; built-in tools bash, read, write, edit, glob, grep, web_fetch, web_search; MCP, Skills, Vaults, Memory Stores, multi-agent, webhooks, schedules, A2A; Ubuntu 22.04 image with Python 3.12, Node.js 20, Go 1.22, Java 21, Ruby 3.3, PHP 8.3, Rust; apt/pip/npm packages, setup and startup scripts; GitHub repo and OSS bucket mounts; CPU, memory and disk not published; SDKs Python, TypeScript, Go, Java; plugin for DeepSeek Harness (QoderAI/dsh-qoder-cloud-agents, MIT, 4 stars); global and China regions |
| Platform feature | Platform feature | Plans: Starter $39/mo (1 agent hosted all month), Pro $69/mo (2 agents, 2,520 credits)Plans: Starter $39/mo (1 agent hosted all month), Pro $69/mo (2 agents, 2,520 credits), Ultimate $169/mo (5 agents, 27,176 credits), Business $499/mo (10 agents, 50,304 credits), Enterprise custom; annual = 10x monthly. 1 credit = $0.001. Extra environments bill by the hour, sized by agent count: 1 agent (Light) 29 credits/h (~$22/mo), 2 agents (Standard) 45 credits/h, 5 agents (Heavy) 71 credits/h, 10 agents (Max) 134 credits/h. CPU/RAM per environment not published. Web search 18 credits; managed LLM usage at model cost + 15%; mailboxes $8-10/mo | Running with no agents: same as running; stopped: ? | An environment bills every hour it runs, with or without agentsAn environment bills every hour it runs, with or without agents; when the credit balance reaches zero its agents pause and the environment stays in place; a Stopped status means powered off; no idle auto-suspend stated | — | Oct 2026 | Managed hosting for OpenClaw, Hermes and Codex agents, aimed at agencies that run agents for clientsManaged hosting for OpenClaw, Hermes and Codex agents, aimed at agencies that run agents for clients; one environment per client with its own keys; schedules, triggers (GitHub, Stripe, cal.com), chat channels (Slack, Telegram, Discord, WhatsApp, Signal and more), integrations and email inboxes; shell in the console; bring your own model key; hard spend caps; CLI, SDK and MCP from Ultimate up; per-client cost tracking and proposals |
| Platform feature | Platform feature | App hosting: Hobby $0 (100K requests, 10M CPU-ms, 1 GB storage per month)App hosting: Hobby $0 (100K requests, 10M CPU-ms, 1 GB storage per month); Builder $7/mo, early access (2M requests, 50M CPU-ms, 10 GB); Enterprise custom. Above the allowance: $0.60 per 1M requests, $0.04 per 1M CPU-ms, $0.03/GB-month. No price published for agent workspaces; AI builds run on your own agent and model, with no markup | $0 compute while idle; storage $0.03/GB-month above the allowance | Dynamic Workers scale to zero; idle agents hibernate to disk and resume in under aDynamic Workers scale to zero; idle agents hibernate to disk and resume in under a second; idle time between requests is not billed | WebAssembly | — | Capability-secured cloud for WebAssembly components (WASI 0.2/0.3, WIT-typed capabilities, q64 languageCapability-secured cloud for WebAssembly components (WASI 0.2/0.3, WIT-typed capabilities, q64 language, Wasmtime 46) and an AI app builder; OpenShell-compatible agent sandboxes run OpenCode, Claude Code or Codex with any model through OpenRouter, on Mac, Windows, Linux or in the browser; qube deploy to *.qubepod.app; runs on the Cloudflare edge, your own hosts (BYOH) or IoT devices; blue-green and canary deploys; MCP to build, deploy and observe; append-only audit journal; help docs are still placeholders |
| Platform feature | Platform feature | Self-hosted free (Apache-2.0); hosted Standard $15/mo ($12/mo yearly) with 25Self-hosted free (Apache-2.0); hosted Standard $15/mo ($12/mo yearly) with 25 computer-hours/month at 4 vCPU / 8 GB; Pro $49/mo with 100 h at 8 vCPU / 16 GB; Max $99/mo with 200 h at 16 vCPU / 32 GB; Rakazo credits from $20/mo (20 h at 4 vCPU / 8 GB; $30 includes 30 h); models on your own Claude, ChatGPT or API key, or credits at provider price with no markup; 7-day free trial with 10 computer-hours | $0 compute when paused (hours count only while in use); storage cost not stated | Rakazo stops or pauses a hosted computer after 10 minutes without useRakazo stops or pauses a hosted computer after 10 minutes without use; included computer hours count only while a bot uses its computer | VM | Aug 2026 | Open-source 'Grok Bot alternative': chat with AI bots (teammates) that each get a computer to run codeOpen-source 'Grok Bot alternative': chat with AI bots (teammates) that each get a computer to run code, edit files and browse; watch the screen and take over; routines in Markdown, schedules, approvals, audit log; web, desktop, iOS and Android apps; integrations and MCP; EU hosting option; site publishes a sandbox price comparison for self-hosters; by Inbox Zero Inc.; GitHub elie222/rakazo 8.4k stars |
| Platform feature | Platform feature | Credits: Agent Computer compute $0.10/hr while activeCredits: Agent Computer compute $0.10/hr while active, $0 when idle (machine size not stated); model calls at cost in credits, or bring your own key or a Claude Max or Gemini subscription; plans Personal $0 (daily free usage), Pro $20/mo, Team $20/seat/mo, Enterprise custom; plans don't change task cost | $0 when idle | Pause saves a full memory and disk snapshot (resume ~100-500 ms, processes continue)Pause saves a full memory and disk snapshot (resume ~100-500 ms, processes continue); hibernate saves disk only (cold boot ~5-7 s, files intact); memory state is kept on local SSD and removed after 8 hours of inactivity, after which the VM cold boots from durable filesystem diffs; Agents API environments are allocated on first use and deleted with their Session unless shared | VM | Mar 2026 | Durable agent harness: OpenAI Agents API-compatible endpoint (api.rebyte.ai/v1), model-agnosticDurable agent harness: OpenAI Agents API-compatible endpoint (api.rebyte.ai/v1), model-agnostic; Rebyte owns the model loop and the sandbox; managed environments with packages, setup commands, vault credentials, skills, MCP and shared environments; Agent Computers in the app run Claude Code, Codex, Gemini CLI and OpenCode; iOS app and Slack; sandbox-sdk (TypeScript, E2B-derived, MIT, 0 stars) with a network-policy API and React/Express lifecycle helpers, no hosted endpoint baked in |
| Platform feature | Platform feature | Per minute by size: small 2 vCPU/8 GB $0.004/min included ($0.24/hr) or $0.006/minPer minute by size: small 2 vCPU/8 GB $0.004/min included ($0.24/hr) or $0.006/min overage ($0.36/hr); large 4 vCPU/16 GB $0.008/$0.012 per min; ultra 8 vCPU/32 GB $0.016/$0.024 per min; Developer $50/mo includes $50 of usage; Team $200+/mo; Enterprise custom; $100 free usage; model costs not included. No 2 vCPU/4 GB size | $0 while sleeping | Workspaces sleep after 30 min without activity (compute pausedWorkspaces sleep after 30 min without activity (compute paused, no charges) and wake on the next message with their branch and agent session; archived after 7 days asleep; deleted 3 days after archiving | VM | Mar 2026 | Cloud coding agents (Claude Code, Codex, Cursor, OpenCode on your own model account) started from Slack, LinearCloud coding agents (Claude Code, Codex, Cursor, OpenCode on your own model account) started from Slack, Linear, GitHub, the dashboard, CLI or API; warm pools and warm hooks; desktop, terminal and Docker inside the workspace; PR review in the app; mobile and desktop apps; SOC 2 Type I; Replicas Group Inc., San Francisco |
| Platform feature | Platform feature | Starter free; Core $20/mo ($18 billed annually) with $20 toward modelsStarter free; Core $20/mo ($18 billed annually) with $20 toward models; Pro $100/mo ($90 annually) with $100 credits, higher tiers to $2,000/mo; Enterprise custom; Reserved VM deployments 0.5 vCPU/2 GB $15/mo, 1 vCPU/4 GB $35/mo, 2 vCPU/8 GB $50/mo, 4 vCPU/16 GB $130/mo; Autoscale $2/mo base + $0.60 per million compute units + $0.40 per million requests; data transfer $0.05/GB. No 2 vCPU / 4 GiB size | Autoscale: no compute charge while idle, plus $2/month base feeAutoscale: no compute charge while idle, plus $2/month base fee; workspace sandbox: ? | Lifecycle of the Agent's cloud sandbox is not stated on the pages readLifecycle of the Agent's cloud sandbox is not stated on the pages read; Autoscale deployments go idle after 15 min without requests and start on the next request; Reserved VM deployments run continuously | VM | Sep 2024 | AI app builder: Replit Agent writes, tests and publishes apps from chatAI app builder: Replit Agent writes, tests and publishes apps from chat; Agent can work on a GitHub repository in an attached cloud sandbox; Postgres and App Storage per app; VNC pane streams a native desktop; Autoscale, Static, Scheduled and Reserved VM deployments; workspaces for teams |
| Platform feature | Platform feature | No published prices ('Book a demo'). Agent computers can run in the customer's own AWSNo published prices ('Book a demo'). Agent computers can run in the customer's own AWS, GCP or Azure account (Terraform or CloudFormation) with Runtime managing the control plane, or the whole platform can be self-hosted with a Helm chart | $0 compute while paused (docs); storage price not stated | Sessions auto-pause after 20 minutes without activity (frozen, no compute chargesSessions auto-pause after 20 minutes without activity (frozen, no compute charges, filesystem kept, resume in seconds); optional ttl_minutes hard limit destroys the session; destroy wipes the filesystem | VM | Jan 2026 | Site now sells an 'AI agent harness for payment teams' (payment ops, risk, fraud, KYC/KYB, compliance, support)Site now sells an 'AI agent harness for payment teams' (payment ops, risk, fraud, KYC/KYB, compliance, support); each session runs Claude Code, Codex, OpenCode, GitHub Copilot, Cursor, Devin or Gemini CLI in an isolated VM with terminal (WebSocket PTY), file API, live HTTPS previews and git; REST API, runtm-api Go CLI, MCP server at mcp.runtm.com; approvals in Slack or Teams; triggers from Slack, Teams, email, SMS, voice, alerts and schedules; SOC 2 Type II; open-source runtm repo (server AGPLv3, CLI and sandbox Apache-2.0, templates MIT; 305 stars) |
| Platform feature | Platform feature | Explore $0 (daily credits, limited access to a Windows cloud computer)Explore $0 (daily credits, limited access to a Windows cloud computer); Pay as you go $50/mo = 50,000 credits ($50 of usage) for AI models and computer time, dedicated Windows or Linux cloud computer billed only while it runs, up to 5 at once, top-ups available; Sai Unlimited $500/mo (unlimited AI usage within fair use, always-on Windows or Linux cloud computer); Enterprise custom. No per-hour computer-time rate or machine size published | $0 computer time when not running (pay-as-you-go plan) | Pay-as-you-go cloud computer is billed only while it runsPay-as-you-go cloud computer is billed only while it runs; Sai Unlimited includes an always-on cloud computer kept available through the end of the billing cycle; Sai can also run on your own Mac or Windows PC | — | Sep 2026 | Computer-use agent: you give a task in plain English and Sai works a real desktop by seeing the screenComputer-use agent: you give a task in plain English and Sai works a real desktop by seeing the screen, clicking and typing; Sai API (POST /v1/agents/message), Python SDK and @simular-ai/sai-mcp MCP server for Claude Code, Codex, Cursor; model-agnostic; replays saved workflows as code (up to 90% fewer tokens); asks a human for sign-ins and credentials via approval links; site claims top OSWorld 2.1 partial score (79.38%, $14.34/task, Oct 2026); built on open-source Agent S framework |
| Platform feature | Platform feature | Flat plans, devboxes included, no metered billing: Jar $49/mo (up to 2 agentsFlat plans, devboxes included, no metered billing: Jar $49/mo (up to 2 agents, 2 devboxes of 2 vCPU/4 GB with 40 GB disk, $5/mo AI credits); Bakery $99/mo (up to 4 agents, 5 devboxes of 4 vCPU/8 GB with 60 GB disk); annual $490/$990; add-on 1 vCPU/4 GB RAM $15/mo; Sokko Memory $25/mo; Enterprise custom; 3-day free trial. Devboxes are not sold separately | Included in flat plan (idle-stopped devbox keeps data, no separate charge) | Devbox lives 24 h by default, extendable up to 72 h at a timeDevbox lives 24 h by default, extendable up to 72 h at a time; idle-stops after 4 h quiet and keeps data, next deploy resumes it; on expiry the machine and data volume are reclaimed (repo secrets kept) | — | — | Devboxes are disposable branch-preview machines that Sokko's hosted agents (OpenClaw, Hermes) deploy toDevboxes are disposable branch-preview machines that Sokko's hosted agents (OpenClaw, Hermes) deploy to; ~16 stacks auto-detected incl. docker-compose, Next.js, Django, Rails; preview URLs on dev.sokko.run behind org sign-in, private via your Tailscale tailnet, or public; share links 5 min-24 h; write-only repo secrets; agents run in US or EU; SOC 2 in progress; operated by Tekiya OU |
SyntixHosted service appears shut down (syntix.pro does not resolve) Platform feature | Platform feature | The hosted service is gone; code free, open source under AGPL-3.0Pricing page | — | Execution machines are stopped after 30 minutes without activity (a cleanup job runsExecution machines are stopped after 30 minutes without activity (a cleanup job runs about every 35 minutes) and are not auto-destroyed (FlyMachineCleanupService.cs, FlyMachineSetup.cs) | Container | Feb 2025 | AI code interpreter chat: the AI writes code, installs dependencies and runs it with full internet accessAI code interpreter chat: the AI writes code, installs dependencies and runs it with full internet access; JavaScript, TypeScript, Python, C#, Java, Go, Rust (the code also has a PostgreSQL executor); upload files and folders up to about 7 GB; custom execution environments from a Dockerfile; unfinished browser extension; machine sizes in code: shared 1 CPU/256 MB or shared 4 CPU/1 GB; syntix.pro no longer resolves in DNS (NXDOMAIN); 7 GitHub stars; last push 2026-02-07 |
| Platform feature | Platform feature | $0.11 | $0 while asleep | A sandbox idle for 3 minutes goes to sleep and uses no box-hours (docs and FAQA sandbox idle for 3 minutes goes to sleep and uses no box-hours (docs and FAQ; the pricing page's rate table says 10 idle minutes); the next message wakes it in a few seconds; a sandbox asleep for a long time moves to storage and takes a few minutes to wake; unpushed work is saved after every turn; archiving a session deletes its sandbox | VM | Apr 2026 | Runs Claude Code, Codex and PostHog Code in parallel cloud sandboxes, one session per taskRuns Claude Code, Codex and PostHog Code in parallel cloud sandboxes, one session per task, each on its own branch and ending in a pull request; fixes failing CI, merge queue, code review, PR dashboard, Slack, Linear and GitHub Issues triggers, scheduled automations; internet on/off switch per session; dev-server port forwarding to localhost in the desktop app; Mac desktop app and web app; REST API and MCP server; open source (MIT), 17 GitHub stars |
| Platform feature | Platform feature | $0.1326 | $0 compute when stopped or snapshotted; storage charges not stated | Billed per running second; stopped and snapshotted VMs accrue no computeBilled per running second; stopped and snapshotted VMs accrue no compute; session sandboxes are ephemeral and destroyed when the session ends; no idle timeout stated | VM | Sep 2026 | Full Linux VM with 60fps desktop, terminal and file explorer shown beside the agent chatFull Linux VM with 60fps desktop, terminal and file explorer shown beside the agent chat; VS Code, Chromium, Ghostty preinstalled; runs Claude Code, Codex, OpenCode, Cursor, Pi; claims 20% cheaper than E2B and Daytona; up to 32 vCPU/128 GiB; self-hosted on AWS/GCP/Azure, VPC or air-gapped for enterprise; public API/SDK/MCP creates Tembo agent sessions, not raw VMs |
| Platform feature | Platform feature | Cloud: $75/seat/month + $0.80 per runner-hour (8 vCPU runnerCloud: $75/seat/month + $0.80 per runner-hour (8 vCPU runner; by default a runner offers one sandbox per vCPU); Hybrid (your own runners): $75/seat/month; Self-hosted: annual license, contact sales; model usage billed by Anthropic to your own API key; no 2 vCPU/4 GiB size | — | One sandbox per task; the sandbox stays up while the agent is idle so you can open it atOne sandbox per task; the sandbox stays up while the agent is idle so you can open it at any time; limits and teardown not stated | Container | — | Board for running a dozen coding-agent tasks at once, each in its own sandbox with the repo clonedBoard for running a dozen coding-agent tasks at once, each in its own sandbox with the repo cloned, services from .tinsmith/config.json started and health-checked, and the dev server on its own port with a public preview URL; chat, diff, terminal and running-app views per task; automations from Sentry, Linear, GitHub, PagerDuty, Datadog, Jira, Zendesk and Snyk with shadow mode and rate caps; agents fix failing checks, answer review comments and rebase; Anthropic models only; GitHub only; hosted control plane with Tinsmith or your own runners, or fully self-hosted |
| Platform feature | Platform feature | No sandbox price from TrueFoundry: sandboxes run on your own Daytona account through yourNo sandbox price from TrueFoundry: sandboxes run on your own Daytona account through your Daytona API key. TrueFoundry plans: Developer $0 (up to 3 users, 10K requests and tool calls per user per month); Pro $25/user/month (20K per user, then $20 per 100K); Enterprise custom (VPC, on-prem, air-gapped) | Set by your own Daytona account | Created only when an agent needs code, files, skills or Code ModeCreated only when an agent needs code, files, skills or Code Mode; reused across turns in the same session (files persist); stopped after idle time, then archived and eventually deleted per the Daytona provider settings (no idle time given) | — | Jul 2026 | Sandbox tool inside TrueForge, TrueFoundry's open-source agent harness (MIT, 6,482 GitHub stars)Sandbox tool inside TrueForge, TrueFoundry's open-source agent harness (MIT, 6,482 GitHub stars), which the TrueFoundry platform embeds unchanged; reusable sandbox environments (base image, build script, CPU/memory/disk, env vars, egress allow list, secrets) baked into Daytona snapshots; default environment 1 CPU, 1 GB RAM, 3 GB disk; egress unrestricted by default, with a Block All Network option; model and MCP credentials stay in the harness; Code Mode and large tool responses use the sandbox; hosted harness available as SaaS, self-hosted or on-prem; HTTP API and TypeScript SDK |
| Platform feature | Platform feature | Free, open source (AGPL-3.0); self-hosted only, and you pay your own model providerPricing page | N/a (self-hosted) | No idle state: each run gets a fresh container that is removed when the run ends (filesNo idle state: each run gets a fresh container that is removed when the run ends (files under /ump/outputs are kept as artifacts); per-run cost, turn and time limits plus a daily workspace budget | Container | Sep 2026 | Self-hosted agent for recurring jobs: you describe a job in plain language, the agent runs it in a sandbox andSelf-hosted agent for recurring jobs: you describe a job in plain language, the agent runs it in a sandbox and, after three matching successful runs, it can graduate to a script that runs without the model; cron, webhook, API and manual triggers; MCP servers (HTTP credentials stay on the host); Anthropic or any OpenAI-compatible model (Ollama, vLLM, OpenRouter); per-job Dockerfile; default sandbox 1 CPU and 1024 MB; Docker or Kubernetes install; 172 GitHub stars |
| Platform feature | Platform feature | No public pricing; invite only (request an invite or book a demo)No public pricing; invite only (request an invite or book a demo); an Enterprise plan adds zero data retention | — | Each task gets its own desktop and its agent keeps working in the background while youEach task gets its own desktop and its agent keeps working in the background while you are away; no suspend, timeout or lifecycle stated | VM | Sep 2026 | An 'agent-native computer' in the browser for law firms, banks, investment firms and consultanciesAn 'agent-native computer' in the browser for law firms, banks, investment firms and consultancies; each task gets its own desktop (Desktops launched 2026-09-22) with its documents, mail and chats; built-in mail, calendar, files, editor, document compare, research, projects, an app store and a Library of skills, agents and apps; Word, Excel, PowerPoint and Outlook add-ins; connects to Outlook, SharePoint, iManage and NetDocuments; agents send nothing without the user's approval; SOC 2 Type II, SAML SSO, GDPR with processing in the EU; Vesence AB (Stockholm and San Francisco); started as 'Cursor for Lawyers'; no terminal, API or SDK mentioned |
| Platform feature | Platform feature | Per seat, launch pricing until 2026-12-31: Team $24.90/seat/month ($19.90 annualPer seat, launch pricing until 2026-12-31: Team $24.90/seat/month ($19.90 annual; $29.90 from 2027-01-01) with $40 LLM + $20 sandbox starting credit and 10 coding sandboxes at once; Business $49.90/seat/month ($39.90 annual) with $60 LLM + $30 sandbox credit, 25 sandboxes at once, BYOK; Enterprise custom (own sandbox vendor account, on-prem); sandbox usage beyond the credit billed on the invoice at unstated rates; 30-day free trial after a demo | Idle sandboxes pause; paused cost not stated | Idle sandboxes pause on their own (pricing page)Idle sandboxes pause on their own (pricing page); after each agent turn Volobox commits and pushes the branch so work survives a stopped sandbox; sessions can be paused, stopped or archived | — | Aug 2026 | Team workspace (projects, chat, docs, meetings, files, support portal) where agents take tasks and open pull requestsTeam workspace (projects, chat, docs, meetings, files, support portal) where agents take tasks and open pull requests; coding sessions in a sandbox per branch with live terminal, diff and branch previews; OpenCode on the organization's models, Cursor and Codex on each person's own account, Claude Code only on a person's own machine; sessions on your own machine; HTTP API and hosted MCP server; cloud, private cloud or on-premise; by Volosoft (ABP Framework) |
| Platform feature | Platform feature | No sandbox price on the site; the harness VM runs on the user's own computerNo sandbox price on the site; the harness VM runs on the user's own computer; Yougori sells inference set-up, megakernel work and GPU brokering | — | — | VM | Sep 2026 | Runs any of 21 coding or security-testing harnesses against the customer's own Yougori inferenceRuns any of 21 coding or security-testing harnesses against the customer's own Yougori inference; works on a copy of a folder and writes back only reviewed files; API keys stay outside the VM; network choice: only the inference, listed sites, or the whole internet; Yougori app and CLI open source under AGPL-3.0-only (10 GitHub stars) |
| Platform feature | Platform feature | Beta, billed by open-platform API usage: model tokens at the GLM-5.3Beta, billed by open-platform API usage: model tokens at the GLM-5.3, GLM-5.3-Flash and GLM-5.3-FlashX inference prices; sandbox and runtime environment free for a limited time until 2026-10-31, later rules 'per platform announcement' (no sandbox rate published); Web Search / Web Fetch tool calls free until 2026-10-31, then charged per call; GLM Coding Plan quota can't be used | Sandbox was free until 2026-10-31; no current rate published | Sessions go idle between turns and resume on a new eventSessions go idle between turns and resume on a new event; each session's sandbox is released when the session ends; /workspace, /tmp and runtime-installed packages are not kept between turns unless the session is created with x-checkpoint: true, which snapshots the system disk and keeps it 1-30 days (default 30); /mnt/session/outputs is always kept and catalogued as downloadable files | Container | May 2026 | Zhipu's hosted agent runtime, the GLM counterpart of Claude Managed Agents: Agents (model, system prompt, tools, MCPZhipu's hosted agent runtime, the GLM counterpart of Claude Managed Agents: Agents (model, system prompt, tools, MCP, Skills), Environments (packages, network policy unrestricted or limited with allowed_hosts, egress only on ports 80/443), Sessions and SSE events; credential vaults the sandbox can't read; memory stores; multi-agent threads; cron deployments; webhooks; Feishu, WeCom, DingTalk and WeChat channels; GitHub repo resources; preinstalled Python, Node, Bun, Go, Rust, Java, Ruby, PHP, C/C++, PostgreSQL, Redis, LibreOffice, TeX, FFmpeg, Playwright with Chromium; docs in Chinese |
| Platform feature | Platform feature | Free during a limited community preview with your own API keys (you pay AnthropicFree during a limited community preview with your own API keys (you pay Anthropic, OpenAI, Google or another provider directly); enterprise commercial engagements on request | — | — | VM | Sep 2026 | Windows desktop AI agent app that operates Windows apps on one or more virtual screens while you keep your own desktopWindows desktop AI agent app that operates Windows apps on one or more virtual screens while you keep your own desktop; watch live and take over; per-action approval settings for files, commands, plugins and MCP tools; review and undo; plugins (Gmail, Slack, GitHub, Google Drive), MCP servers, scheduled tasks, custom skills; Windows only |
| Platform feature | Platform feature | Free trial $0 for 14 days; Basic $18/mo (2 vCPU / 8 GB RAM, $10/mo AI included)Free trial $0 for 14 days; Basic $18/mo (2 vCPU / 8 GB RAM, $10/mo AI included); Pro $64/mo (4 vCPU / 16 GB, $40/mo AI); Ultra $200/mo (8 vCPU / 32 GB, $100/mo AI); 100 GB disk on every plan; media models billed per request. No 2 vCPU / 4 GiB plan, and plan prices include AI credits | Flat monthly plan; paid computers do not sleep | Free trial: computer sleeps when idle and ordinary access ends after 14 days (files keptFree trial: computer sleeps when idle and ordinary access ends after 14 days (files kept; up to seven one-hour recovery sessions); paid plans run always on | — | Jun 2025 | Personal cloud computer running the Zo agent 24/7; root SSH for IDEs; public API (/zo/ask) and MCP serverPersonal cloud computer running the Zo agent 24/7; root SSH for IDEs; public API (/zo/ask) and MCP server; hosts Sites, Services, custom domains and a zo.space page; talk to it by SMS, email, Discord, Slack or Telegram; use Claude Code or Codex subscriptions or bring your own LLM key; Hermes or OpenClaw can run inside; regular filesystem snapshots, files kept in an open container image format; desktop app |
| Browser infra | Browser infra | Free $0 (1,000 credits plus a 10,000 one-time bonus, 3 sessions)Free $0 (1,000 credits plus a 10,000 one-time bonus, 3 sessions); Starter $29/mo (from 30,000 credits, 3 sessions); Professional $189/mo (from 225,000 credits, 30 sessions); Enterprise $558/mo (from 775,000 credits, 100 sessions); custom; 10% off annually; credits per session-hour not stated on the pricing page | $0 (idle sessions are terminated) | Browser sessions end after 10 idle minutes by default (timeoutMinutes configurable)Browser sessions end after 10 idle minutes by default (timeoutMinutes configurable); billed in 30 s increments; saved profiles keep cookies and local storage between sessions | — | Nov 2024 | Cloud browsers for agents via API and SDK; Agent Builder compiles described automations into deterministic codeCloud browsers for agents via API and SDK; Agent Builder compiles described automations into deterministic code; 'Mark' agent; integrated or custom proxy; integrations with n8n, Zapier, Make, Claude Code and Codex; HIPAA and SOC 2 Type 2; site banner: 'Airtop is working on something new' |
| Browser infra | Browser infra | Infrastructure: $0.01 per browser created + $0.05 per browser-hour (to the nearestInfrastructure: $0.01 per browser created + $0.05 per browser-hour (to the nearest minute) + Anchor proxy $8/GB (own proxy or no proxy $0.20/GB egress) + AI steps $0.01-$0.10 each; Automation Task runs $0.10 each, or the infrastructure cost if higher; plans Free $5 credits/month, Starter $50/mo, Growth $2,000/mo, overage $1 per credit, Enterprise custom; the homepage also lists a Team plan at $500/mo and its cost calculator uses $0.09/browser-hour; vCPU and memory not stated | No paused state; an idle session bills browser-hours until idle_timeout ends it | No pause; a session ends after idle_timeout with no connections (default 5 minNo pause; a session ends after idle_timeout with no connections (default 5 min; -1 disables) or at max_duration (default 180 min, no upper limit); the VM is destroyed at session end; persistent volumes and profiles carry files and logins across sessions | VM | — | Headful Chromium controlled over CDP (Playwright, Puppeteer), REST or OS-level inputHeadful Chromium controlled over CDP (Playwright, Puppeteer), REST or OS-level input; natural-language web tasks and versioned Automation Tasks; managed authentication with identities and MFA (OmniConnect), 1Password; persistent volumes and profiles; recordings, live view, batch sessions, webhooks; residential proxies, Anchor VPN, captcha solving, Cloudflare Web Bot Auth; code execution on Growth and up; regions US, plus EU (Frankfurt), Asia and Australia on Growth; bring-your-own-cloud or on-prem on Enterprise; SOC 2 Type 2, ISO 27001, GDPR; BAA/HIPAA on Enterprise |
| Browser infra | Browser infra | Free, open source (MIT); self-hosted (benchmarks run on an AWS m5zn.metal host)Free, open source (MIT); self-hosted (benchmarks run on an AWS m5zn.metal host); no hosted service or prices | N/a (self-hosted) | Browser sessions created from a warm pool (microVM snapshots) and released after useBrowser sessions created from a warm pool (microVM snapshots) and released after use; no hosted lifecycle | VM | May 2026 | Open-source browser hosting control plane and host runtime for Chromium sessions over CDPOpen-source browser hosting control plane and host runtime for Chromium sessions over CDP; claims 224 ms p50 full lifecycle (create + CDP connect + page.goto + release) on a self-hosted BrowserArena-method run, 2026-06-01; warm pools; benchmark harness and run artifacts in the repo; one-person project by Vividh Mahajan; 37 GitHub stars |
| Browser infra | Browser infra | Billed by traffic: the docs say the Browser API spends 5 credits per MBBilled by traffic: the docs say the Browser API spends 5 credits per MB; free tier 5,000 credits a month (about $7.50 value, nearly 1 GB of Browser API traffic), pre-paid wallet; pay-as-you-go $/GB and plan prices are on brightdata.com, which could not be loaded; no per-hour or vCPU/GiB rate | — | — | Browser | — | Remote managed browsers for Puppeteer, Playwright and Selenium (wss://brd.superproxy.io:9222)Remote managed browsers for Puppeteer, Playwright and Selenium (wss://brd.superproxy.io:9222), part of Bright Data's Unlocker Scraping Suite; automatic proxy rotation including residential IPs, fingerprint management, CAPTCHA detection and solving, retries and session recovery; unlimited concurrent sessions; session ID and session logs API; Bright Data MCP server, CLI and agent skills for coding agents; agents can register an account themselves and get a pre-created agent_browser_api zone |
| Browser infra | Browser infra | $0.06 per browser-hour; home-IP proxy traffic $5/GB; your own proxy or direct $0.20/GB$0.06 per browser-hour; home-IP proxy traffic $5/GB; your own proxy or direct $0.20/GB; CAPTCHA solving included; pay as you go with non-expiring credits, $1 free; agent tasks extra (bu-ultrafast $0.006/task, model tokens); enterprise custom | Same as running ($0.06/browser-hour until stopped) | A browser runs and bills until stopped through the API (closing CDP does not stop it)A browser runs and bills until stopped through the API (closing CDP does not stop it); agent-run browsers stop about 2 minutes after the last run, keepAlive 120-14,400 s; idle agent browsers are billed as browser time | — | Oct 2024 | Stealth browsers driven over CDP with Playwright, Puppeteer or Selenium; 0.6 s median readyStealth browsers driven over CDP with Playwright, Puppeteer or Selenium; 0.6 s median ready; residential proxies by country; live takeover for 2FA; recording; concurrency 10 free up to 1,000 by total paid; on-prem option; SOC 2 Type II, HIPAA, ZDR; open-source browser-use library (MIT, 121k stars) |
| Browser infra | Browser infra | Free and self-hosted (README says MIT; the repo has no LICENSE file)Free and self-hosted (README says MIT; the repo has no LICENSE file); you pay for your own server and LLM keys (price per hour for 2 vCPU / 4 GiB: n/a (self-hosted)) | N/a (self-hosted) | Browser sessions and saved profiles persist until closed or until the container restartsBrowser sessions and saved profiles persist until closed or until the container restarts; the dev branch backs profiles up to Cloudflare R2 every 5 minutes and on session close | Container | Nov 2025 | REST API compatible with the Browser-Use Cloud API v2 spec; live VNC view at /vnc.htmlREST API compatible with the Browser-Use Cloud API v2 spec; live VNC view at /vnc.html; reusable browser sessions and profiles with saved logins; task start, stop, pause and resume with step updates; custom HTTP tools the agent can call; 15+ LLMs (GPT, Claude, Gemini, Groq, Llama); multi-arch Docker image reqeique/browser-use-api; MCP support 'coming soon'; 8 GitHub stars; last push 2025-12-21 |
| Browser infra | Browser infra | Compute by browser size and session time: Compute M 2 vCPU / 2 GB $0.06/hrCompute by browser size and session time: Compute M 2 vCPU / 2 GB $0.06/hr, Compute L 4 vCPU / 4 GB $0.12/hr (no 2 vCPU / 4 GB size); proxy traffic datacenter $1/GB, residential $6/GB, mobile $8/GB, premium residential $12/GB, BYOP fee $1/GB; CAPTCHA $0.01 standard, $0.10 premium per clear; plans Free $0 (1 concurrent browser, datacenter proxy only), Hobby $5, Developer $10, Startup $50, Pro $100, Business $500 per month, each fully converted to usage credits; Enterprise custom | — | Sessions are created and deleted through the API and billed by session timeSessions are created and deleted through the API and billed by session time; the DELETE response returns runtime seconds and stored cookies; no idle timeout or maximum session length found | Container | May 2026 | Stealth cloud browser API for AI agents, scraping and automationStealth cloud browser API for AI agents, scraping and automation; Playwright connect, REST ('Humanized REST' with a Browser Flow JSON schema and action catalog) and MCP browser server; sessions ready in about 100 ms (site claim); stealth on by default against Cloudflare, DataDome and Google; custom fingerprints and full identity control; zero-logs policy (page content, screenshots, cookies, keystrokes not retained); agent skill and setup templates for coding and personal agents; Agent Actions coming soon; pays in card, Bitcoin, Ethereum, USDC; the site is also served at browserciity-website.pages.dev |
| Browser infra | Browser infra | Browser-hours, billed per minute with a 1-minute minimum per session: Free $0/mo (1Browser-hours, billed per minute with a 1-minute minimum per session: Free $0/mo (1 browser-hour, 3 concurrent, 15 min/session); Developer $20/mo with 100 h then $0.12/browser-hour, 25 concurrent; Startup $99/mo with 500 h then $0.10/browser-hour, 100 concurrent; Scale custom; proxies $12/GB (Developer) or $10/GB (Startup) after included GB; Fetch API $1 or $0.50 per 1k calls; Search API from $5/1k; vCPU and memory per browser not stated | No paused state; a kept-alive session bills browser time until released orNo paused state; a kept-alive session bills browser time until released or timed out | No pause; a session ends at disconnect unless keepAlive is set (paid plans)No pause; a session ends at disconnect unless keepAlive is set (paid plans), and every session ends at its project or per-session timeout, maximum 6 h; keepAlive sessions run until released or timed out; Contexts save cookies and auth across sessions; Functions invocations 60-900 s with no persistent storage | VM | Jun 2024 | Cloud browsers driven with Playwright, Puppeteer, Selenium or Stagehand; Search and Fetch APIsCloud browsers driven with Playwright, Puppeteer, Selenium or Stagehand; Search and Fetch APIs; Agent Identity (Cloudflare-verified agents, captcha solving, proxies); Functions run TypeScript next to a browser (us-west-2); live view, session replay, logs; regions US East, US West, EU Central, UK (APAC on request for enterprise); no on-prem; SOC 2 Type II, HIPAA (BAA); claims 35M+ sessions a month for 10,000 customers; Stagehand framework open source (MIT, 26k stars) |
| Browser infra | Browser infra | Monthly credit pools: Starter $29 (500,000 credits, 5 concurrent sessions)Monthly credit pools: Starter $29 (500,000 credits, 5 concurrent sessions); Pro $99 (2,000,000 credits, 50 concurrent); Scale $299 (10,000,000 credits, 100 concurrent); runtime 1 credit per second, datacenter proxy 35 credits/MB, residential proxy 350 credits/MB, AI selectors 15 credits per 1,000 tokens, CAPTCHA 75 credits per solve; 5,000 free trial credits, no card; unused credits expire monthly; no vCPU/GiB sizes | $0 after a session closes or expires (credits stop when the session is$0 after a session closes or expires (credits stop when the session is destroyed) | Sessions are destroyed when closed and auto-expire at their timeout (default 300 sSessions are destroyed when closed and auto-expire at their timeout (default 300 s, 10 to 1800 s); plan caps of 15 min (Starter), 30 min (Pro) and 1 hour (Scale) per session; page content not stored beyond the session | Browser | Apr 2026 | REST browser API for AI agents that returns markdown, stable element refs (e1, e2), form metadataREST browser API for AI agents that returns markdown, stable element refs (e1, e2), form metadata, page stability signal and change diffs instead of raw HTML; steps for observe, click, fill_form, extract with AI selectors and declarative schemas, scroll-collect, JavaScript, screenshots and PDF; auto-dismiss cookie banners; datacenter and residential proxies; auto CAPTCHA solving; Python, TypeScript and Ruby SDKs, OpenAPI 3.1 spec, MCP server for Cursor, Claude Desktop, Claude Code and Windsurf |
| Browser infra | Browser infra | Self-hosted licenses: $119/seat/year commercialSelf-hosted licenses: $119/seat/year commercial, $49/seat/year non-commercial (a seat is one concurrent user); 7-day paid evaluation ($20 per seat with a work email, $40 otherwise, credited on upgrade, up to 10 seats); BrowserBox + KRNL bundle $179/seat/year. Hosted Cloud API minutes, prepaid and never expiring: $5 for 60 min (about $0.08/min), $20 for 300 min (about $0.07/min), $60 for 1,000 min (about $0.06/min); no vCPU or memory size stated | Cloud API minutes are reserved per sessionCloud API minutes are reserved per session; unused minutes return when you delete it | Cloud API sessions are created for a set number of minutes (up to 120-minute sessions onCloud API sessions are created for a set number of minutes (up to 120-minute sessions on Professional); unused minutes return when you delete the session; Fleet allocations are ephemeral; a Persistent Browser (always-on profile, $29/mo) is 'coming soon' | Browser | Nov 2019 | 60 FPS streamed browser with no client install; embed a live session with the <hyper-frame> element and drive tabs60 FPS streamed browser with no client install; embed a live session with the <hyper-frame> element and drive tabs, navigation and capture from JavaScript; policies for clipboard, uploads, downloads, printing and developer tools, with audit logs; scripted automation with human takeover for CAPTCHAs and logins; session recording; runs on Windows, macOS, Linux and LXC, with or without GPU; Docker deployment; reach sessions over Cloudflare tunnels, Tor, ZeroTier or HTTPS; KRNL text-mode browser over SSH; GitHub BrowserBox/BrowserBox (3,934 stars) under its own commercial license, not open source; repo describes it as browser infrastructure for regulated enterprises, AI and agentic microtasks |
| Browser infra | Browser infra | Credit plans: Free $0 (1,000 credits, 1 concurrent, 1 browser-hour)Credit plans: Free $0 (1,000 credits, 1 concurrent, 1 browser-hour); Hobby $19/mo (120,000 credits, 5 concurrent, 120 browser-hours); Startup $99/mo (1M credits, 50 concurrent, 1,000 h); Business $249/mo (3M credits, 100 concurrent, 3,000 h); pricing page says 1 browser-hour = 1,000 credits (about $0.08-$0.16 per browser-hour on paid plans), but the docs' credit table says 1 credit per browser-second; /content 1 credit, /pdf and /screenshot 5 credits; proxies 12 credits/MB datacenter, 80 (pricing page) or 110 (docs) credits/MB residential; flat-rate Dedicated Clusters $690/mo for 100 parallel sessions up to $1,250/mo for 400; vCPU and memory not stated | No paused state; an open browser uses credits until it closes or times out | No pause; a browser runs until the client closes it or the session timeout set at launchNo pause; a browser runs until the client closes it or the session timeout set at launch (Free plan up to 15 min per session; paid plan caps not stated); contexts and cache can persist user data and site cache across sessions | Container | 2020 | Hosted headless or headed Chrome for Puppeteer and Playwright over WebSocket, plus HTTP APIs (content, PDFHosted headless or headed Chrome for Puppeteer and Playwright over WebSocket, plus HTTP APIs (content, PDF, screenshot); spin-up under 300 ms claim; 1,000+ concurrent sessions; stealth mode, fingerprint randomization, captcha solving; datacenter and residential proxies with geotargeting; live view and session recording; n8n, Zapier, Make integrations; self-funded; by BrowserCloud, LLC |
| Browser infra | Browser infra | Routing on your own provider keys free up to 10,000 sessions/monthRouting on your own provider keys free up to 10,000 sessions/month, then $0.0005 per session; session recording $0.007 per video minute; observability broadcast $0.05 per 1,000 events; managed browsers draw from a prepaid wallet (rate not shown); top-ups from $5; provider bills are separate; Enterprise custom | — | Holds the browser when a client connection drops so you can reconnect to the same pageHolds the browser when a client connection drops so you can reconnect to the same page; saved login profiles load into any provider's browser and save back at session end; otherwise follows the provider's lifecycle | Browser | Mar 2026 | 'Open router for cloud browsers': one CDP endpoint (wss://cdp.browsergateway.io) with failover across providers and'Open router for cloud browsers': one CDP endpoint (wss://cdp.browsergateway.io) with failover across providers and routing strategies such as a priority order; up to 100 saved login profiles, encrypted at rest; session replay exported as MP4; hosted MCP server and REST API (screenshot, content, scrape); works with Playwright, Puppeteer, Stagehand, browser-use, Claude Code, Codex; self-hostable npm package (browser-gateway); MIT license, 19 GitHub stars; by Monostellar Labs |
| Browser infra | Browser infra | Units of up to 30 s of browser time per connection (120 units per browser-hourUnits of up to 30 s of browser time per connection (120 units per browser-hour; each reconnect starts a new unit): Free $0 (1k units/mo, 2 concurrent); Prototyping $25/mo billed yearly (20k units, $0.0020/unit overage); Starter $140/mo yearly (180k units, $0.0017/unit); Scale $350/mo yearly (500k units, $0.0015/unit); prices shown are billed yearly, and the site says yearly billing saves 30% (monthly prices not in the page source); overage is $0.18-$0.24 per browser-hour; residential proxy 6 units/MB, datacenter 2 units/MB, captcha 10 units; Enterprise custom with self-hosting; vCPU and memory not stated | No paused state; a reconnect-held browser keeps consuming units until itsNo paused state; a reconnect-held browser keeps consuming units until its timeout | No pause; a session ends when the client disconnects unless reconnect is usedNo pause; a session ends when the client disconnects unless reconnect is used, which keeps the browser alive server-side for a set timeout; max session time by plan (Free 2 min, Prototyping 15 min, Starter 30 min, Scale 60 min); saved sessions and replays kept 1-90 days by plan; saved login profiles | VM | Nov 2017 | Hosted Chrome, Chromium, Firefox and WebKit over CDP/WebSocket for Puppeteer and PlaywrightHosted Chrome, Chromium, Firefox and WebKit over CDP/WebSocket for Puppeteer and Playwright; BrowserQL, REST APIs (screenshot, PDF, content, scrape), BAP SDK (TypeScript, Python), MCP server and managed agent API; stealth, captcha solving, residential proxies; live URL hybrid automation; endpoints in San Francisco, London and Amsterdam; self-host via Docker or Kubernetes with GPU and air-gapped options; open-source repo under SSPL or commercial license (13.9k stars) |
| Browser infra | Browser infra | Free, open source (MIT); runs on your own Kubernetes cluster (the README claims up to 85%Free, open source (MIT); runs on your own Kubernetes cluster (the README claims up to 85% lower cost on EKS at scale than managed services such as Browserbase) | N/a (self-hosted) | No persistence: a browser runs until deleted with DELETE /browsers/{id}No persistence: a browser runs until deleted with DELETE /browsers/{id}; browser persistence and session state are on the roadmap | Container | Jul 2025 | Self-hosted open-source alternative to Browserbase: REST API to launch, listSelf-hosted open-source alternative to Browserbase: REST API to launch, list, inspect and delete sandboxed Chrome instances, and a CDP WebSocket proxy through a FastAPI head node; deploy on AWS, GCP, Azure, on-prem or a local Kind cluster; Helm chart and Terraform for AWS EKS (AKS and GKE planned); API-key auth; examples for Browser Use, LangChain and CrewAI; 165 GitHub stars |
| Browser infra | Browser infra | $5 per GB of egress (decimal GB), prepaid, residential proxies included$5 per GB of egress (decimal GB), prepaid, residential proxies included; no per-session, idle or seat fees; $10 minimum top-up; 100 MB free on signup; credits never expire; no per-hour or vCPU/GiB rate | $0 (no idle-session charges; billing is per GB transferred) | Sessions auto-expire after ttl_seconds (default 1800 s, 60 to 3600 s)Sessions auto-expire after ttl_seconds (default 1800 s, 60 to 3600 s); provisioning takes about 15 to 25 s; up to 10 concurrent sessions per workspace | Browser | Apr 2026 | Stealth browser API for agents and scrapers: POST /v1/sessions returns a CDP WebSocket URL for Playwright, PatchrightStealth browser API for agents and scrapers: POST /v1/sessions returns a CDP WebSocket URL for Playwright, Patchright, Puppeteer or raw CDP; fingerprints drawn per session from a corpus of real browser captures (default Chrome on Android); per-session JA4 TLS/HTTP2 signature; residential egress in 12 regions from the same company's Marvel proxy network; screenshots and live view; captcha handling guide; releases verified against FingerprintJS Pro; docs MCP server; made by HYPR Pte. Ltd., Singapore |
| Browser infra | Browser infra | Free $0: 1 browser-hour/month, 1 running browserFree $0: 1 browser-hour/month, 1 running browser; Basic $20/mo: 250 hours, 10 concurrent, then $0.08/hour; Pro $99/mo: 1,250 hours, 100 concurrent, then $0.06/hour; runtime metered per second; the browser's vCPU and memory are not stated | $0 runtime while paused; storage allowances and retention of paused browsers$0 runtime while paused; storage allowances and retention of paused browsers are "still being finalized" | Auto-pause after 5 minutes with no connected agent or Live View (on by defaultAuto-pause after 5 minutes with no connected agent or Live View (on by default, can be turned off); pause closes Chrome and keeps the profile (cookies, local storage, files) and links but not in-memory page state; resume restarts Chrome; paused browsers take no concurrency slot; the Free plan pauses running browsers at its hour limit | Browser | Nov 2026 | Persistent Chrome browsers for agents with saved login statePersistent Chrome browsers for agents with saved login state; pause, resume and fork (a fork snapshots the paused browser's filesystem into a new browser); persistent CDP links for Playwright, Puppeteer and agent-browser; Live View for people, read-only or read & write, shareable; REST API with OpenAPI spec and a Go CLI; works with Claude Code, Codex, Cursor and OpenClaw; separate from CloudBrowser AI (cloudbrowser.ai, the CloudBrowser.ai row) |
| Browser infra | Browser infra | Basic $25/month: 250 browser hours (30,000 usage units), 10 concurrent instancesBasic $25/month: 250 browser hours (30,000 usage units), 10 concurrent instances, 3 tabs each, 7-day free trial; Premium $90/month: 1,000 browser hours (120,000 units), 25 concurrent instances; Custom: contact; one usage unit = 30 s of browser time per tab; annual plans 2 months free; premium proxies and live remote desktop on every plan | — | Browser closes automatically after keepOpen seconds of inactivity (default 300 sBrowser closes automatically after keepOpen seconds of inactivity (default 300 s; 0 = never close); optional saveSession saves cookies and localStorage (not open tabs) under a label, and recoverSession restores them in a later browser | Container | Oct 2024 | Real stealth Chrome in the cloud for ChatGPT, Claude and custom agentsReal stealth Chrome in the cloud for ChatGPT, Claude and custom agents; hosted MCP server at mcp.cloudbrowser.ai with 20 browser tools (OAuth or API token; connector open source, MIT); HTTP Browser API, then CDP with Puppeteer or Playwright; official .NET and Node.js SDKs plus toolkits for OpenAI Agents SDK and Claude Agent SDK; skills for Claude Code, Codex and Gemini; human-like fingerprints, CAPTCHA detection, live remote desktop for human takeover; AI extraction endpoints (query, summarize, translate, describe image); n8n node; ChatGPT plugin in review; 14-day money-back guarantee |
| Browser infra | Browser infra | Workers Free: 10 browser-minutes per day, 3 concurrent browsersWorkers Free: 10 browser-minutes per day, 3 concurrent browsers; Workers Paid: 10 browser-hours/month included, then $0.09 per browser-hour, plus 10 concurrent browsers included (monthly average of daily peaks), then $2.00 per extra browser (Browser Sessions only); usage summed daily in seconds and rounded to the nearest hour per month; vCPU and memory not stated | No paused state; an idle browser counts browser time until the inactivityNo paused state; an idle browser counts browser time until the inactivity timeout (60 s default, up to 20 min) closes it | No pause; a browser closes after 60 s of inactivity by defaultNo pause; a browser closes after 60 s of inactivity by default, extendable with keep_alive up to 20 minutes; no fixed maximum lifetime while active; sessions also close when Cloudflare rolls out a new release; sessions can be reused | V8 isolate | May 2023 | Quick Actions REST endpoints (markdown, screenshot, PDF, snapshot, links, scrape, JSON via promptQuick Actions REST endpoints (markdown, screenshot, PDF, snapshot, links, scrape, JSON via prompt, crawl) and Browser Sessions with Puppeteer, Playwright, Stagehand or raw CDP from any environment; Playwright MCP and CDP for MCP clients such as Claude Code and Cursor; up to 200 concurrent browsers per account on Workers Paid; runs on Cloudflare's global network |
| Browser infra | Browser infra | Billed on bandwidth, not time: Browsers $5/GB of traffic, Web Data API $3 per 1k resultsBilled on bandwidth, not time: Browsers $5/GB of traffic, Web Data API $3 per 1k results; Start $0 ($5 credits, 5 concurrent browsers, up to 1 GB), Grow $49/mo ($49 credits, 30 concurrent, up to 10 GB), Scale $129/mo ($129 credits, 100 concurrent, up to 26 GB), Custom; dedicated IPs $5/address/month. The older browser.cash/developers page still lists Browser as a service at $0.09/hr and Browser Agents at $0.5/M input and $2/M output tokens, with a 1-hour session limit | $0 for a session that moves no bytes$0 for a session that moves no bytes, but it holds a concurrency slot until it stops | No pause: a session ends when you stop it, at its duration (60 to 3,600 s, default 3,600No pause: a session ends when you stop it, at its duration (60 to 3,600 s, default 3,600, no extension) or if its machine goes away; persisted profiles carry state into a new session; warm browsers in a pool aren't billed until acquired, and a lease timeout caps how long one is held | VM | Jul 2025 | Hosted Chrome on real devices on ISP connections in the country you pickHosted Chrome on real devices on ISP connections in the country you pick, reachable over CDP a few seconds after creation; guides for Patchright, Playwright, Puppeteer, Browser Use, Stagehand and Crawl4AI; browser pools, live view, profiles, proxies, CAPTCHA solving and ad blocking, extensions, native input; Web Data API, managed data feeds and an ISP network; Browser Cash also runs a browser extension that rewards people with points for letting their browsers serve as nodes; open-source tools in the BrowserCash GitHub org (teracrawl, MIT, 290 stars) |
| Browser infra | Browser infra | Free, open source (AGPL-3.0); self-hosted with no usage feesFree, open source (AGPL-3.0); self-hosted with no usage fees; you pay only for your own server (price per hour for 2 vCPU / 4 GiB: n/a (self-hosted)) | N/a (self-hosted) | A session runs until you or the agent stop itA session runs until you or the agent stop it; logged-in sessions are saved in vaults encrypted with your password and resumed later; after an update, saved sessions restart once you enter the password | Container | Oct 2026 | Real Google Chrome that Claude Code, Codex or any agent drives over a built-in MCP server or REST, one token per agentReal Google Chrome that Claude Code, Codex or any agent drives over a built-in MCP server or REST, one token per agent; human takeover for logins, 2FA and CAPTCHAs, with Slack or Discord webhooks when an agent needs you; own Rust compositor streams lossless tiles plus AV1 over UDP instead of VNC (claims 55 fps at 1.8 Mbit/s vs 12.4 Mbit/s for KasmVNC); per-session proxy; session timelines and screen recordings; uploads and downloads with approval; installer sizes parallel sessions to server memory; admin reachable only through an SSH tunnel, Tailscale or Cloudflare Tunnel; 11 languages; 87 GitHub stars |
| Browser infra | Browser infra | Credits: 2 per browser minute with Playwright code, 7 per minute with a promptCredits: 2 per browser minute with Playwright code, 7 per minute with a prompt, +2 per minute for zero data retention; 1-minute minimum; plans Free 1,000 credits/mo, Hobby $19/mo (5,000), Standard $99/mo (100,000), Growth $399/mo (500,000), Scale $749/mo (1M), less ~16-20% billed yearly; extra credits $5 per 1,000 (Hobby) to 5,000 (Scale); so one code-driven browser-hour (120 credits) is about $0.12-$0.60 in extra credits; vCPU and memory not stated | No paused state; an idle session bills browser minutes until its activityTtl orNo paused state; an idle session bills browser minutes until its activityTtl or ttl ends it | Session destroyed after activityTtl seconds of inactivity (10-3600Session destroyed after activityTtl seconds of inactivity (10-3600, default 300) or at its total ttl (30-3600 s; docs table says default 600, API schema says 300); no pause; persistent profiles save cookies and localStorage on close, but files on disk are lost | Container | Feb 2026 | Remote browser with Playwright and agent-browser CLI (60+ commands) preinstalledRemote browser with Playwright and agent-browser CLI (60+ commands) preinstalled; run Python, JavaScript or bash in the sandbox; CDP URL for your own Playwright; interactive live view; persistent profiles; country proxies; API, Node and Python SDKs, Vercel AI SDK; up to 20 concurrent sessions on all plans at launch; docs now point CLI and MCP agents to Interact (scrape + interact) and call the CLI browser command legacy; Firecrawl is YC-backed and open source (201.8k stars shown on site) |
| Browser infra | Browser infra | Annual contracts in euros: standalone Glazer Browser from €2,499/year (annual seatAnnual contracts in euros: standalone Glazer Browser from €2,499/year (annual seat subscription; a seat is one concurrent user or one API-controlled browser instance); Agent packages from €9,999/year include 1 to 5 Browser seats plus Federation data credits; additional seats available | — | Persistent or disposable sessions; zero-retention sessionsPersistent or disposable sessions; zero-retention sessions; personas (sock-puppet identities) persist across sessions | Container | 2025 | OSINT and investigations platform for law enforcement, intelligence, compliance and journalismOSINT and investigations platform for law enforcement, intelligence, compliance and journalism; isolated remote browser for open, deep and dark web with built-in VPN and Tor, managed personas, automatic archiving of visited content with cryptographically timestamped screenshots, real-time entity enrichment; API and AI-agent integration for API-controlled browser instances; Glazer Agent and Federation data products; SaaS or on-premises; sandbox start cut from 8-10 s to about 2 s (blog); Estonian company (Glazer Technologies OÜ), part of the EU co-funded Cyber Accelerator |
| Browser infra | Browser infra | Browser hours included in Gologin plans, then extra hours at $0.09 (Professional)Browser hours included in Gologin plans, then extra hours at $0.09 (Professional), $0.08 (Business), $0.07 (Enterprise) or $0.05 (Custom) per browser-hour; extra concurrent sessions $2/month each; proxy traffic beyond the included 2 GB $1/GB; Free: 3 runs of up to 15 minutes on 1 session. The pricing docs list Professional $9/mo ($4.5/mo billed annually) with 50 h and 1 session, Business $99/mo ($49.5) with 100 h and 2, Enterprise $179/mo ($89.5) with 200 h and 3, Custom from $279/mo ($139.5) with 400 h and 4; the cloud browser page instead shows $4.50, $8 and $14 a month; vCPU and memory per browser not stated | — | Profiles keep cookies, storage, device fingerprint and IP between runsProfiles keep cookies, storage, device fingerprint and IP between runs; a cloud browser stays active for 15 minutes after the last action, or after you close its tab without pressing Stop; the comparison table says paid plans have no session length limit | Browser | — | Real Chromium (Gologin's Orbita) in the cloud, reached over CDP from PlaywrightReal Chromium (Gologin's Orbita) in the cloud, reached over CDP from Playwright, Puppeteer or Python by changing one connect line; persistent profiles keep the device fingerprint, cookies and residential or mobile IP, so logins survive between runs; Live View to watch or take over for 2FA; Gologin Agent Browser CLI and MCP server for AI agents; Web Unlocker, Cloud Android and Gologin Proxies on the same account; REST API at 300 to 1,200 requests a minute by plan |
| Browser infra | Browser infra | Compute hours: Free $0/mo (3 h, $5 trial AI credit)Compute hours: Free $0/mo (3 h, $5 trial AI credit); Developer $25/mo (100 h, then $0.12/h; 100 job runs then $0.12/run); Startup $120/mo (500 h, then $0.10/h; 400 job runs then $0.10/run); Large machines +$0.18/h (Developer) or $0.15/h (Startup), X-Large +$0.48/h or $0.40/h; Enterprise volume pricing; machine sizes standard 6 shared vCPU/2 GB, large 8 shared vCPU/4 GB, x-large 1 performance vCPU/8 GB, so no 2 vCPU/4 GiB size | — | No pause; automations run as API Runs (default 10-minute timeoutNo pause; automations run as API Runs (default 10-minute timeout, extendable with extendTimeout) or scheduled Jobs on a pool that keeps minCount machines ready and scales to maxCount; AuthSessions keep logins between runs; human-in-the-loop can pause a Run for input or hand over the browser | — | Jun 2026 | Code-first browser automation platform: Intuned Agent writes and fixes Playwright code (TypeScript or Python) from aCode-first browser automation platform: Intuned Agent writes and fixes Playwright code (TypeScript or Python) from a prompt and schema, or copies a site's network calls; deploys automations as APIs and scheduled Jobs with webhooks; self-healing scrapers; AuthSessions for logged-in integrations; captcha solving, advanced stealth and proxies; human-in-the-loop handoff; traces and observability; browser-use integration; SSO on Enterprise |
| Browser infra | Browser infra | Per second: headless browser $0.0000166667 ($0.06/hr); headful $0.0001333336 ($0.48/hr)Per second: headless browser $0.0000166667 ($0.06/hr); headful $0.0001333336 ($0.48/hr); headful + GPU $0.0008000016 ($2.88/hr); app invocation $0.0000666668/s (4 GB at $0.0000166667/GB-s); plans: Developer free + usage with $5 credits/mo, Hobbyist $30/mo + usage with $10 credits, Start-Up $200/mo + usage with $50 credits, Enterprise custom | $0 in standby (concurrency slot still held) | Browsers enter standby after 10 s with no CDP, WebDriver BiDi, Live ViewBrowsers enter standby after 10 s with no CDP, WebDriver BiDi, Live View, telemetry or computer-controls activity; standby keeps browser state at zero usage cost (but holds the concurrency slot); timeout countdown starts in standby (timeout_seconds default 60, max 72 h); GPU browsers don't support standby; profiles preserve browser state between sessions | Browser | Feb 2025 | Browser infra for agents: cloud Chromium (headless or headful) in under 30 ms from hot poolsBrowser infra for agents: cloud Chromium (headless or headful) in under 30 ms from hot pools, under 1 s cold (site claims); CDP, WebDriver BiDi, Playwright execution inside the browser VM, computer controls for screenshot agents; stealth with captcha solving and proxies; managed auth; profiles; MCP server, CLI, OpenAPI; concurrency Developer 5, Hobbyist 10, Start-Up 150 browsers; SOC 2 and HIPAA; open source kernel-images (Apache-2.0, 1,135 GitHub stars); public notice of a security incident (compromised API key, logging-provider data accessed) on 2026-11-04 |
| Browser infra | Browser infra | Agent plans by concurrency: Free $0 (3 concurrent sessions, US exit)Agent plans by concurrency: Free $0 (3 concurrent sessions, US exit); Pro $500/mo (list price $1,000; 25 concurrent sessions, every location); Scale $3,000/mo (100 concurrent sessions); Enterprise custom; unlimited bandwidth on every plan; hosted browser sessions need an eligible paid plan. Chrome extension for people (Ghost and Shield) $10/mo or $100/yr, 30-day free trial | — | Session ends when the agent ends it or the browser tab closesSession ends when the agent ends it or the browser tab closes; retention not stated (the site tells users to verify retention themselves) | Browser | Sep 2026 | Web access for agents: routing through Legba exits to reach public pages, bot walls and users' logged-in accountsWeb access for agents: routing through Legba exits to reach public pages, bot walls and users' logged-in accounts, plus hosted isolated browser sessions; installed as an agent skill (npx skills add legba-app/skills; MIT, 13 stars), with OpenClaw listed as supported; integrations with existing browser frameworks; Chrome extension with Ghost (private route) and Shield (remote isolated browser); Legba, Inc. |
| Browser infra | Browser infra | Monthly plans paid one month at a time: Free $0 (10 browser hours, 3 concurrentMonthly plans paid one month at a time: Free $0 (10 browser hours, 3 concurrent, 1,000 Web Fetch requests, 30 MB proxy); Pro $12 (100 hours, 25 concurrent, 10 profiles, 1 GB proxy, 100 CAPTCHA solves); Pro+ $60 (500 hours, 100 concurrent); Enterprise $180 (1,500 hours, 250 concurrent); extra Web Fetch $0.51 to $0.39 per 1k requests and extra proxy traffic $10.00 to $8.00/GB by plan; no browser-hour overage rate stated | — | A session runs until you stop it (closing the Playwright or Puppeteer connection doesn'tA session runs until you stop it (closing the Playwright or Puppeteer connection doesn't end it); persistent Contexts keep browser state between sessions; no idle timeout or maximum duration stated | Browser | Sep 2025 | Cloud Browser API over CDP for Playwright and Puppeteer, plus Web Fetch, Web Search and a browser AgentCloud Browser API over CDP for Playwright and Puppeteer, plus Web Fetch, Web Search and a browser Agent; the Moli engine starts faster and uses less memory than Chrome (site claim) and is open source (github.com/lexmount/moli, Apache-2.0, 23,754 stars); Chromium for extensions, persistent contexts and replay; official or custom proxies; CAPTCHA processing; live view and session records; Runtime runs JavaScript tasks in the cloud with connected browsers and a local client (access enabled per project); CLI and agent Skill; Node.js and Python SDKs |
| Browser infra | Browser infra | Explorer $0 forever (10 browser-hours/month, 5 concurrent)Explorer $0 forever (10 browser-hours/month, 5 concurrent); Builder $19/mo (300 browser-hours, then $0.08/browser-hour, 30 concurrent); Enterprise custom with on-prem or private cloud; bring your own proxy; vCPU and memory not stated | No paused state; an open session counts browser time until it closes | No pause; browser time is metered per second while a session is openNo pause; browser time is metered per second while a session is open; a CDP connection closes after 15 minutes regardless of plan; an MCP session closes after 5 minutes without a request | — | Feb 2023 | Headless browser engine written from scratch in Zig for automation and AI agents (JavaScript executionHeadless browser engine written from scratch in Zig for automation and AI agents (JavaScript execution, no graphical rendering); claims 9x faster and 16x less memory than Chrome in its benchmark; CDP server for Puppeteer, Playwright and chromedp, WebDriver BiDi, MCP server, HTTP API, built-in agent and Python SDK; cloud regions west Europe and west US; Chrome also offered in the cloud; Lightpanda 1.0 released; open source (AGPL-3.0, 42k stars) |
| Browser infra | Browser infra | Browser sessions $0.01/hour; Browser Functions $0.05/hourBrowser sessions $0.01/hour; Browser Functions $0.05/hour; residential proxies $10/GB (your own proxy $0.20/GB); LLM tokens at cost; Free $0 with $10 one-time credits (5 concurrent, 15-min sessions); Developer $20/mo with $20 credits (25 concurrent, 120 min); Startup $100/mo with $100 credits (100 concurrent, 300 min); Enterprise custom with on-prem; vCPU and memory not stated | No paused state; an idle session bills browser time ($0.01/h) until its idleNo paused state; an idle session bills browser time ($0.01/h) until its idle timeout closes it | Session closes after idle_timeout_minutes without an operation (default 3Session closes after idle_timeout_minutes without an operation (default 3, max 30) or at max_duration_minutes (default 15, docs max 1,440; plan caps 15/120/300 min); no pause; saved state through vaults, personas and authenticated profiles | Browser | Dec 2024 | Cloud browser sessions over CDP for Playwright plus an AI SDK (act, choose and check page steps) and web agentsCloud browser sessions over CDP for Playwright plus an AI SDK (act, choose and check page steps) and web agents; Functions deploy browser automations as serverless API endpoints with cron schedules; stealth, captcha solving, residential proxies; secret proxies fill cards, passwords and 2FA without Notte seeing them; live view and recordings; open source (SSPL, 2.1k stars); claims 20,000 developers |
| Browser infra | Browser infra | Engine free and open source (Apache-2.0), self-hosted in one Docker containerEngine free and open source (Apache-2.0), self-hosted in one Docker container; managed Obscura Cloud only through a demo (the README calls it a waitlist); no published prices | — | Sessions boot on demand and are torn down when done | Browser | Apr 2026 | Headless browser engine in Rust for agents and scrapingHeadless browser engine in Rust for agents and scraping, a drop-in for headless Chrome over CDP with Puppeteer and Playwright; built-in anti-detection; claims 30 MB memory, a binary of about 70 MiB, 85 ms page loads and session starts under 50 ms; native screenshots, screencasts and PDF export without Chromium; the README says Obscura inspired the first prototype of Cloudflare's Kitesurf browser; 35,423 GitHub stars (the site shows 16.2k) |
| Browser infra | Browser infra | Cloud Developer $49.99/mo (cloud sandbox for evaluation: 1 seat, 5 active contextsCloud Developer $49.99/mo (cloud sandbox for evaluation: 1 seat, 5 active contexts, 3 VM profiles); 14-day trial $0.99 one-time; Self-Hosted Starter $1,999/mo (unlimited contexts); Business $19,999 one-time + $3,999/mo support; Enterprise $49,999 one-time + $9,999/mo support; self-hosted plans are flat licenses with no per-minute billing | — | — | Browser | Feb 2026 | Stealth browser engine by Olib AI: fingerprint controls compiled into C++ (canvas, WebGL, audio, fonts)Stealth browser engine by Olib AI: fingerprint controls compiled into C++ (canvas, WebGL, audio, fonts); up to 256 contexts per instance with under 12 ms cold start per context; Agent Rendering (OwlMark) structured page view instead of screenshots, claimed ~10x fewer tokens; 186 automation tools; MCP server built in at POST /mcp; built-in Tor per context; local Qwen3-VL-2B vision model for CAPTCHA solving on self-hosted tiers; RFC 9421 HTTP Message Signatures for Cloudflare Verified Bots (Show HN); Node and Python SDKs and CLI; free Owl Light browser; "SOC 2 aligned, not certified" |
| Browser infra | Browser infra | Billed per GB of traffic from a shared Oxylabs credit balance: $8/GB on the free welcomeBilled per GB of traffic from a shared Oxylabs credit balance: $8/GB on the free welcome credits ($5, up to 0.6 GB) and Starter ($15/mo, up to 2 GB); $7.5/GB on Basic ($100/mo, up to 13 GB); $5.5/GB on Flexible ($500/mo, up to 90 GB); $4/GB on Enterprise (from $2,500/mo, 625+ GB); no time-based price | Billed per GB of traffic; an open session holds a concurrency slot | No pause: a session lasts until the client closes itNo pause: a session lasts until the client closes it; an unclosed session stays open, holding one of the 100 concurrent sessions, until it is reclaimed; persistent profiles (default cap 5) and persistent sessions (default 5 concurrent) keep state | Browser | — | Remote browsers over WSS/CDP for Playwright and Puppeteer, with Oxylabs residential proxies on every sessionRemote browsers over WSS/CDP for Playwright and Puppeteer, with Oxylabs residential proxies on every session; automatic and manually triggered CAPTCHA solving (hCaptcha, reCAPTCHA, Turnstile); country and device targeting; session inspection over VNC and recording; MCP for AI agents; 10 new sessions a second, 100 concurrent; GitHub oxylabs/agent-browser is a docs and examples repo (665 stars, no license) |
| Browser infra | Browser infra | Desktop app free with 500 agent steps a month; Oya Cloud Free: 1 cloud browser hourDesktop app free with 500 agent steps a month; Oya Cloud Free: 1 cloud browser hour; Developer $20/mo (100 browser hours, 1 GB proxy, 5,000 agent steps); Startup $99/mo (500 hours, 5 GB, 50,000 steps); beyond that $0.10 or $0.08 per browser hour, $8/GB and $0.002 per step; self-hosting free up to 5 concurrent cloud browsers; no vCPU or memory size stated | — | Durable session state and persistent personas with cookies and proxiesDurable session state and persistent personas with cookies and proxies; no idle timeout stated | Container | Mar 2026 | Browser control plane for healthcare and revenue-cycle work: an agent does a portal task onceBrowser control plane for healthcare and revenue-cycle work: an agent does a portal task once, Oya saves it as Playwright code (a playbook) and replays it with new inputs without a model; hash-chained audit trail; HIPAA compliant, signs BAAs, SOC 2 Type II audit in progress; live view and human takeover; roles and scoped credentials; sealed passwords and TOTP secrets; signed lifecycle webhooks; MCP, TypeScript SDK and CLI; self-host with Docker, ECS, Kubernetes or Google Cloud; source on GitHub under the Sustainable Use License (SDK and CLI MIT), 349 stars |
| Browser infra | Browser infra | Hosted: x402 API $0.01 per 5-minute block (USDC on Base, no signup)Hosted: x402 API $0.01 per 5-minute block (USDC on Base, no signup); MCP: 1 credit = 10 minutes, $5 buys 100 credits; Enterprise dedicated capacity by contact; self-hosted: free, open source (MIT); browser vCPU and memory not stated | N/a (sessions are deleted, not paused) | No pause: one fresh browser per task, removed when the session is deleted or its TTLNo pause: one fresh browser per task, removed when the session is deleted or its TTL expires (ttlSeconds per session, capped by an operator maximum; a TTL controller cleans up); x402 sessions are bought in 300-second blocks | Browser | Aug 2026 | Headful Chromium for one task, controlled over CDP (Playwright) with a live view a person can take overHeadful Chromium for one task, controlled over CDP (Playwright) with a live view a person can take over; remote MCP server for Claude Code and Codex with OAuth 2.1 and card top-ups; attestation proofs that a session ran on genuine confidential hardware with approved images; warm pools per region, 470 ms p50 / 889 ms p95 allocation and 102,859 sessions in 30 days (launch post); browser image based on Kernel's open-source kernel-images; Helm charts, local Kind workflow and GKE guide; GitHub reclaimprotocol/popcorn-oss 10 stars; built for Reclaim's identity-proof flows |
| Browser infra | Browser infra | Hosted: a one-time free allowance of 48 browser hours with 5 browsers at a time (noHosted: a one-time free allowance of 48 browser hours with 5 browsers at a time (no card); larger plans priced by sales on concurrency and hours; proxy traffic billed separately at your Rayobyte residential rate (residential proxies from $0.0095/GB); self-hosted rayobrowse free and unlimited | Same as running (an open session uses browser hours until closed or expired) | No pause: a session keeps running after the client disconnectsNo pause: a session keeps running after the client disconnects, so an agent can reconnect with its cookies and tabs; it ends when closed, when the browser dies, or at its lifetime (default 7,200 s), and every second counts as browser time | Browser | Jan 2026 | Stealth Chromium for scraping and AI agents, reached over CDP from Playwright, Puppeteer, Selenium or ScrapyStealth Chromium for scraping and AI agents, reached over CDP from Playwright, Puppeteer, Selenium or Scrapy; a residential proxy with a fresh sticky session on every launch; setup prompt for Claude Code, Cursor and Codex; self-hosted version is a Docker image (GitHub rayobyte-data/rayobrowse, MIT, 216 stars); Rayobyte also sells proxies, a Scraping API and a Web Unblocker |
| Browser infra | Browser infra | Active browser-hours: Free $0/mo (1 browser-hour, 1 active browser, 1 profile)Active browser-hours: Free $0/mo (1 browser-hour, 1 active browser, 1 profile); Developer $0.10 first month then $4.90/mo (30 browser-hours, 10 active browsers, 20 profiles); one-time top-up $4.90 for 20 browser-hours ($0.245/h); vCPU and memory not stated | No paused state; browser-hours count while a browser is alive until the idleNo paused state; browser-hours count while a browser is alive until the idle timeout closes it | No pause; an idle-timeout controller watches the CDP heartbeat and closes cold browsersNo pause; an idle-timeout controller watches the CDP heartbeat and closes cold browsers automatically (timeout value not stated); persistent profiles keep cookies, local storage, tabs and login state across sessions | Container | Nov 2026 | Hosted Chromium sessions over CDP for Playwright, Puppeteer, Selenium and browser-useHosted Chromium sessions over CDP for Playwright, Puppeteer, Selenium and browser-use; hardened Chromium with anti-detect fingerprints and banner blocking; residential proxies by country or your own proxy; live viewer with human takeover for 2FA; recordings; Hermes, OpenClaw and coding-agent skill integrations; pages marked 'V0'; by Remote-browser, Inc. (site links the OnFinality-io GitHub org) |
| Browser infra | Browser infra | Library free (package.json says MIT; the repo has no LICENSE file)Library free (package.json says MIT; the repo has no LICENSE file). No price of its own: the Sandbox runs on your Vercel account and is billed at Vercel Sandbox rates (the Vercel Sandbox row) | Vercel Sandbox billing; a stopped or expired Sandbox runs no browser | AgentBrowser.create() starts a fresh Sandbox and close() stops itAgentBrowser.create() starts a fresh Sandbox and close() stops it; otherwise the Sandbox expires at its wall-clock timeout; keepalive() renews the timeout (heartbeat at half the window, capped at 5 min); named sessions (AgentBrowser.session({id})) find the same runtime from another process, and if an expired Sandbox resumes, Chromium starts fresh and page, cookies and tabs are lost; destroy() removes it permanently | Browser | Jul 2026 | TypeScript client (npm remote-agent-browser, v1.3.0) that runs Vercel Labs' agent-browser CLI in a Vercel SandboxTypeScript client (npm remote-agent-browser, v1.3.0) that runs Vercel Labs' agent-browser CLI in a Vercel Sandbox; exec/run agent-browser commands, snapshots with element refs, screenshots, uploads, downloads, traces, HAR and video returned as bytes; proxy option; auth through VERCEL_OIDC_TOKEN; browser image published to Vercel Container Registry; benchmark: 100 pages in 8.01 s in a fresh 2-vCPU Sandbox, 461.6 MB peak; GitHub vercel-labs/remote-agent-browser 185 stars |
| Browser infra | Browser infra | $0.09 per browser hour plus $1.80/GB residential proxy traffic (base rate$0.09 per browser hour plus $1.80/GB residential proxy traffic (base rate, discounted by prepaid subscription tier; tier prices aren't in the page HTML); CAPTCHA solving included; free tier 1 concurrent session and 1 browser hour per month; Enterprise custom | Billed per browser hour; sessions close at their TTL | Session lasts sessionTTL seconds (default 180 s, settable from 60 sSession lasts sessionTTL seconds (default 180 s, settable from 60 s; 900 s is the recommended maximum but longer can be set), then expires and the browser is closed; the product page says sessions auto-release when idle; saved Profiles keep cookies and login state across runs | Browser | Dec 2024 | Cloud browser for scraping and AI agents over a CDP WebSocket (wss://browser.scrapeless.com) for Puppeteer, PlaywrightCloud browser for scraping and AI agents over a CDP WebSocket (wss://browser.scrapeless.com) for Puppeteer, Playwright, Selenium, Stagehand, Browser-Use, Crawl4AI, LangChain, CrewAI and n8n; MCP server (scrapeless-mcp-server) for Claude Code, Cursor and Cline; claims sub-2 s cold starts (P95), thousands of concurrent sessions and 12+ data centers; 70M+ residential IPs in 195 countries; engine-level fingerprint control; CAPTCHA solving; live view and session replay; real-time 2FA signaling for a human hand-off; custom proxies for subscribers only; Node.js and Go SDKs |
| Browser infra | Browser infra | Credits: session time 1 credit per 30 s (120 credits per browser hour) plus bandwidthCredits: session time 1 credit per 30 s (120 credits per browser hour) plus bandwidth (datacenter from 7 credits/MB, residential from 52 credits/MB, cached 1 credit/MB), minimum 5 credits per session; live VNC or WebRTC add-on 5 credits per session; plans Discovery $30/month (200,000 credits, about $0.018 per browser hour), Pro $100/month (1,000,000 credits, $0.012/hr), Startup $250/month ($0.012/hr), Enterprise $500/month ($0.0109/hr); 1,000 free credits on signup; crashed sessions and failed connects free; no vCPU/GiB sizes | — | Session lasts until disconnect by default (auto_close=true)Session lasts until disconnect by default (auto_close=true); with a session ID and auto_close=false the browser keeps running after disconnect and can be resumed with cookies, localStorage, sessionStorage, IndexedDB and proxy assignment; maximum session duration via timeout, default 900 s, maximum 1800 s | Browser | Apr 2026 | Playwright, Puppeteer, Selenium and any CDP client over wss://browser.scrapfly.ioPlaywright, Puppeteer, Selenium and any CDP client over wss://browser.scrapfly.io; Scrapium stealth Chromium (550+ patched source files, real device fingerprints); residential and datacenter proxies in 190+ countries; CAPTCHA solver; resource stubbing and caching to cut bandwidth; session recording and replay; custom Chrome extensions; works with browser-use, Stagehand, LangChain, LlamaIndex, CrewAI; hosted MCP server with Cloud Browser tools; AI Browser Agent product; SDKs for Python, TypeScript, Go, Rust, Scrapy; part of Scrapfly's web scraping platform |
| Browser infra | Browser infra | Credit-based, $0.005 per agent step (1 credit = $0.01)Credit-based, $0.005 per agent step (1 credit = $0.01); Free $0 with welcome credits ($5 per docs), 2 concurrent browsers; Starter $50/month incl. $50 credits, 10 concurrent browsers (docs table says 5); Growth $500/month incl. $500 credits, 50 concurrent browsers; Enterprise custom (single tenant, BYOC, on-prem); no per-hour or per-vCPU rate | — | Sessions are opened and closed by the SDKSessions are opened and closed by the SDK; persistent sessions and browser profiles keep authentication across sessions; no idle timeout or maximum session length found | Browser | Aug 2025 | Cloud browsers plus a web agent built on small language models specialized for web tasks (claims 5x fasterCloud browsers plus a web agent built on small language models specialized for web tasks (claims 5x faster, 7x cheaper); agent and browser co-located; run tasks from natural language, schema-validated JSON extraction, scrape to markdown, screenshots or PDF; stealth Chromium fork, humanized movements, managed proxies, CAPTCHA solver; P2P tunnel to localhost and private networks; live view and video recording; custom tools and extensions; zero data retention option; Smooth CLI for Claude Code, OpenClaw and other agents; Python and TypeScript SDKs (GitHub circlemind-ai/smooth-sdk, 19 stars); 99.99% SLA, SOC 2 aligned, BAA |
| Browser infra | Browser infra | Browser-hours billed per minute, rounded up: Launch $0/mo + usage with $30 one-timeBrowser-hours billed per minute, rounded up: Launch $0/mo + usage with $30 one-time credits, $0.10/browser-hour, 10 concurrent; Scale $250/mo + usage with $100/mo credits, $0.08/browser-hour, 100 concurrent; Enterprise custom; proxy bandwidth $10/GB or $6/GB; CAPTCHA solves $3/1k or $1/1k; /scrape $5/1k; dedicated IPs $5 each (Scale); vCPU and memory not stated | No paused state; an idle session bills browser minutes until its timeout orNo paused state; an idle session bills browser minutes until its timeout or inactivityTimeout releases it | No pause; a session stays alive 5 min by defaultNo pause; a session stays alive 5 min by default, or up to the plan's maximum (Launch 15 min, Scale 1 h, Enterprise 24 h), then is released; optional inactivityTimeout releases it early when idle; profiles keep cookies, auth, extensions and settings across sessions | — | Nov 2024 | Browser API for agents over CDP (Puppeteer, Playwright and other CDP tools); sub-second session start claimBrowser API for agents over CDP (Puppeteer, Playwright and other CDP tools); sub-second session start claim; stealth, CAPTCHA solving, residential proxies; persistent profiles, files and extensions APIs, credentials; live session viewer with human takeover; session replays; Steel Outpost self-hosting in your own cloud; open-source steel-browser (Apache-2.0, 8.0k stars); HIPAA BAA on Scale; by Nen Labs, Inc. |
| Browser infra | Browser infra | Browser-hours billed per minute, rounded upBrowser-hours billed per minute, rounded up; monthly fee becomes usage credit: Pay as you go $0/mo at $0.35/browser-hour; Starter $29/mo at $0.30/h (5 GB traffic); Base $99/mo at $0.25/h (15 GB); Advanced $199/mo at $0.20/h (50 GB); Custom from $399/mo; premium residential and mobile proxies $10-$7/GB; Reserved dedicated fleet from $0.03/browser-hour; vCPU and memory not stated | No paused state; an idle session bills browser minutes until its inactivityNo paused state; an idle session bills browser minutes until its inactivity timeout stops it | No pause; a session runs until stopped or until its inactivity timeout (default 30 sNo pause; a session runs until stopped or until its inactivity timeout (default 30 s, set with inactive_kill_timeout), with no maximum session time; persistent profiles save chosen browser state (cookies, passwords, open tabs) at stop for the next start | Container | — | Anti-detect cloud Chrome over CDP (Playwright, PuppeteerAnti-detect cloud Chrome over CDP (Playwright, Puppeteer, chromedp) with fingerprints from 2.5M+ real devices and 100M+ residential IPs; bypass claims for Cloudflare, DataDome, Akamai and others; CAPTCHA solving on every plan; human emulation; persistent profiles and extensions; shared HTTP cache across sessions; Agent Browser product for AI agents; 8 to 1,000+ concurrent sessions |
| Browser infra | Browser infra | No Agent Browser price on the product page, docs or telnyx.com/pricingNo Agent Browser price on the product page, docs or telnyx.com/pricing; developer preview, request access | — | Short-lived sessions: lifetime 900 s by default (60 to 1800 s)Short-lived sessions: lifetime 900 s by default (60 to 1800 s), closed after 300 s idle by default (30 s minimum); expired sessions end and no cookies, credentials or state carry over between sessions | Container | Oct 2026 | REST API POST /v2/browser_sessions; per-session allowed_domains (1 to 50 patterns)REST API POST /v2/browser_sessions; per-session allowed_domains (1 to 50 patterns); observe returns the accessibility tree with short-lived element refs, no raw HTML or CDP; fixed action set (navigate, observe, click, fill, select, press, scroll, wait_for, screenshot, plus batch of up to 50); screenshots as signed artifacts; idempotency keys; asynchronous provisioning; concurrent-session quota per organization; no CAPTCHA solving, anti-bot evasion or fingerprint randomization; one Telnyx API key also covers Web Search and Inference |
| Browser infra | Browser infra | Free $0 (100 minutes of 2-parallel automation, valid for lifetime)Free $0 (100 minutes of 2-parallel automation, valid for lifetime); Browser Cloud plan $29/month billed annually (included hours and parallel sessions not published; concurrency depends on plan); Enterprise contact sales | — | Session released when the agent is done; 5-minute default session timeoutSession released when the agent is done; 5-minute default session timeout, configurable at creation, maximum not published (per TestMu's own Browserbase comparison citing its session lifecycle docs); optional session persistence carries cookies, local storage and login state to later sessions | — | Mar 2026 | Real cloud Chrome for AI agents via npm @testmuai/browser-cloud SDK, MCP server (testmu-browser-agent) and agent skillReal cloud Chrome for AI agents via npm @testmuai/browser-cloud SDK, MCP server (testmu-browser-agent) and agent skill; Playwright, Puppeteer and Selenium connect to the same sessions; built-in TestMu AI Tunnel to localhost, staging and VPN-only apps; best-effort Stealth Mode (fingerprint masking, ad blocking); Chrome extensions; video replay, console and network logs, command replay; cookbooks for OpenAI, Claude and Gemini computer use; SOC 2 Type II, ISO 27001/27017/27701; built on the LambdaTest browser testing grid |
| Browser infra | Browser infra | Cloud browsers: waitlist, no published price. Fortress engine: free for developersCloud browsers: waitlist, no published price. Fortress engine: free for developers; production use needs a signed key at $2,000/month (fortress.tilion.com/pricing) | — | — | Browser | Jul 2026 | Fortress, the shipped product, is a stealth Chromium fork (Chromium 153) with 81+ C++ fingerprint patches and recordedFortress, the shipped product, is a stealth Chromium fork (Chromium 153) with 81+ C++ fingerprint patches and recorded human mouse paths; Playwright and Puppeteer connect over CDP on localhost:9222; integrations for browser-use, Crawl4AI, Stagehand and LangChain; MCP server with 29 tools for Claude, Codex, Cursor, Windsurf and Cline; Docker, pip, npm, .deb or single binary for Windows, macOS and Linux (plus arm64 Linux); claims it loaded 86.4% of 92 bot-protected sites vs 74.5% for Camoufox; Fortress Source Available License 1.1 (earlier releases BSD-3-Clause); GitHub tiliondev/fortress 802 stars; hosted cloud browsers are in pilots |
| Browser infra | Browser infra | TinyBrowser $0.002 per minute ($0.12/browser-hour), 5 concurrent sessions to startTinyBrowser $0.002 per minute ($0.12/browser-hour), 5 concurrent sessions to start, paid from a prepaid Wallet ($10 minimum deposit; new accounts get $8); TinyAgent $0.016/step; TinySearch and TinyFetch free; Enterprise custom. AgentQL's own pricing page still lists Starter $0/mo (10 remote-browser hours, then $0.12/h, 5 sessions) and Professional $99/mo (500 hours, then $0.10/h, 100 sessions); TinyFish calls credit plans legacy | Billed per browser minute; an idle session runs until its inactivity timeout | Session ends after an inactivity timeout (no CDP commands): timeout_seconds 5 to 86,400Session ends after an inactivity timeout (no CDP commands): timeout_seconds 5 to 86,400, defaulting to the plan maximum (15 min on free plans, 60 min on paid); sessions keep no cookies or storage, but Browser Context Profiles save cookies, local and session storage for later runs; or DELETE the session | Browser | Aug 2024 | TinyFish's web stack for agents: TinyBrowser (cloud Chromium over CDP for Playwright and PuppeteerTinyFish's web stack for agents: TinyBrowser (cloud Chromium over CDP for Playwright and Puppeteer, with residential proxies, sticky exit IP and anti-bot handling), TinyAgent (goal-driven web automation, 89.9% on Online-Mind2Web claimed), TinySearch, TinyFetch, Mako model; session creation takes 10-30 s; live browser preview stream; TinyVault credentials; AgentQL query language with Python and JavaScript SDKs (GitHub tinyfish-io/agentql, MIT, 1,507 stars); Enterprise adds VPC deployment, SSO, audit logs, ISO 27001 |
| Browser infra | Browser infra | Free, open source (MIT); self-hosted on your own Linux Docker host (price per hour for 2Free, open source (MIT); self-hosted on your own Linux Docker host (price per hour for 2 vCPU / 4 GiB: n/a (self-hosted)) | N/a (self-hosted) | Pause and resume per sandbox; the on-disk browser profile and shared files survivePause and resume per sandbox; the on-disk browser profile and shared files survive pause/resume and are removed on delete; website logins are not guaranteed to stay valid | Container | Mar 2026 | Self-hosted headed Chromium for agents such as OpenClaw and Codex: CDP for Playwright and Puppeteer through ticketedSelf-hosted headed Chromium for agents such as OpenClaw and Codex: CDP for Playwright and Puppeteer through ticketed WebSocket URLs, plus SDK, CLI and REST API; live view and human takeover through noVNC or Xpra; admin dashboard; persistent profiles; optional CloakBrowser and Camoufox engines; agent skills for deploying and using it; Python/FastAPI and React; runtime images include GPL Xpra; 28 GitHub stars; last push 2026-11-05 |
| Browser infra | Browser infra | Spark $18/month (12,500 virtual minutes, 2 concurrent channels)Spark $18/month (12,500 virtual minutes, 2 concurrent channels); Forge $495/month (250,000 virtual minutes, $0.002 per extra minute, 30 channels); Foundry custom, billed per successful interaction; 14-day free trial; prices exclude VAT | — | — | Container | Apr 2025 | Augmented web proxy (the site links Surfly's terms and careers pages and cites 12+ years of R&D): agents see the DOMAugmented web proxy (the site links Surfly's terms and careers pages and cites 12+ years of R&D): agents see the DOM and act (click, scroll, fill forms) in a live web session the user also sees, through MCP and an Automation API for extension code; Webfuse stresses it runs in the user's own browser, not a remote headless browser; session recording, audit logs, element masking, access controls, human handoff, co-browsing; SOC 2 Type II and ISO 27001; public cloud, custom regions, optional private cloud or on-premise; NICE CXone integration |
| Browser infra | Browser infra | Credit plans shared by all Zenrows products: Free $0 (5,000 credits/month)Credit plans shared by all Zenrows products: Free $0 (5,000 credits/month); Build $16/mo (45K credits); Launch $58/mo (250K); Growth $166/mo (1.2M); Scale $458/mo (5M); Enterprise custom; the Browser draws 5 credits a minute and residential bandwidth 25,000 credits per GB; failed requests not charged | Billed per browser minute until the session closes | Session TTL default 3 min (180 s), settable from 1 to 15 min maximumSession TTL default 3 min (180 s), settable from 1 to 15 min maximum; the session closes when the TTL ends; to reuse a login you save cookies and local storage yourself (Playwright storage_state, or the MCP tools browser_get_cookies / browser_set_cookies) | Browser | — | Formerly Scraping Browser; cloud browser driven over WebSocket/CDP with your own Puppeteer or Playwright codeFormerly Scraping Browser; cloud browser driven over WebSocket/CDP with your own Puppeteer or Playwright code; residential proxies, country and world-region targeting; anti-bot and CAPTCHA handling; MCP server (mcp.zenrows.com) with browser tools; CLI; Python, Node and Go SDKs (MIT); agent storefront where agents can buy credits; 5 to 1,000+ concurrent requests by plan; SOC 2 and ISO 27001 on Enterprise |
| Computer-use desktop | Computer-use desktop | Sandbox.service.ab0t.com and its /pricing and /faq pages returned HTTP 502 on every trySandbox.service.ab0t.com and its /pricing and /faq pages returned HTTP 502 on every try: curl and WebFetch, then curl again later the same day; no Wayback Machine copy | — | ab0t.com says 'persistent or ephemeral' and 'reset on demand' | — | — | One of eight layers of the ab0t platform (Identity, Permission Aware Proxy, Sandboxes, AuditDB, Events, LLM GatewayOne of eight layers of the ab0t platform (Identity, Permission Aware Proxy, Sandboxes, AuditDB, Events, LLM Gateway, Payments, Agent Harness); cloud desktops and cloud browsers for agents; frame-level recording of every cursor move, click and keystroke; boots in <8 s; per-workspace network egress controls; cloud or on-prem; based in Auckland, NZ |
| Computer-use desktop | Computer-use desktop | $0.08 per workstation minute ($4.80/h; workstation size not stated)$0.08 per workstation minute ($4.80/h; workstation size not stated); Free $0 plus usage (30 workstation hours a month, $100 one-time credit); Pro $50/mo plus usage; Ultimate $200/mo plus usage; early-access signups get $50 in credit; Enterprise custom | — | Workstations come from a pool of ready environments and are released by APIWorkstations come from a pool of ready environments and are released by API (ws.Release()); billed per minute of a requested workstation | — | Apr 2024 | API for agents to request a virtual workstation; browser automation, code execution (Python, Go, Node.js)API for agents to request a virtual workstation; browser automation, code execution (Python, Go, Node.js), Zoom meetings and screen sharing, voice (speak, listen, ask), screen recording, RFB/VNC, AI prompt workflows; Go SDK; custom tools deployed with Kubernetes manifests; homepage copyright 2024; docs.agentstation.ai failed TLS on 2026-11-14 |
| Computer-use desktop | Computer-use desktop | $299 per Runner/month flat: a dedicated Mac with up to 6 parallel iOS simulators (the$299 per Runner/month flat: a dedicated Mac with up to 6 parallel iOS simulators (the launch blog says 4 isolated sessions), unlimited simulator time, no hourly billing; self-serve checkout up to 4 Runners, larger fleets contact sales; no vCPU/GiB sizes published | Flat monthly price; a Runner costs the same idle or busy | A session holds one machine; the client daemon releases it after 15 minutes idleA session holds one machine; the client daemon releases it after 15 minutes idle (configurable; server idle limit 16 minutes) and the server releases it after 24 hours on one machine; on release credentials are revoked, simulators erased and the session volume rebuilt from a clean image, so nothing survives a session | — | Oct 2026 | Cloud iOS Simulators for AI agents and CI on Linux; simulators already booted (no cold start)Cloud iOS Simulators for AI agents and CI on Linux; simulators already booted (no cold start); simctl passthrough (drop-in for xcrun simctl); taps, gestures, typing, screenshots and live preview in a web UI; virtual camera feeds an image or video; two-way port tunnels so local dev servers (e.g. Metro) and hot reload work; works with Argent (MCP server, CLI, Node.js SDK), Maestro, Appium, GitHub Actions; 1 GiB upload limit; screen, keystrokes and app content not recorded, machine logs kept 14 days; activation up to 2 business days after payment; cloud builds and Android emulators announced as coming |
| Computer-use desktop | Computer-use desktop | Prepaid credit billed per minute: tiny (1/2 vCPU, 4 GB RAM, 8 GB disk) $0.07/hPrepaid credit billed per minute: tiny (1/2 vCPU, 4 GB RAM, 8 GB disk) $0.07/h; small (1 vCPU, 6 GB, 12 GB) $0.11/h; medium (2 vCPU, 8 GB, 16 GB) $0.16/h; large (4 vCPU, 12 GB, 20 GB) $0.26/h; no 2 vCPU / 4 GB size. Stopped (snapshot) $0.001/h. VPN add-on: proxy +$0.03/h and $10/GB; unlock +$0.06/h, $15/GB, $3 per 1k connections. Card with auto-refill required to run machines; $1 welcome credit | $0.001/h per stopped machine (snapshot kept) | Runs until stopped or until its ttlMinutes deadlineRuns until stopped or until its ttlMinutes deadline, then onExpire stop (keeps state) or destroy; ttlMinutes null means always on. Stopping snapshots the whole filesystem plus open windows and tabs (memory is not kept, programs relaunch); a save point is also taken every 30 min while running. When credit runs out machines stop and keep state. Idle warm-pool machines bill at running rates | Container | Oct 2026 | Each machine is a small Linux desktop with a persistent browser (multiple profiles)Each machine is a small Linux desktop with a persistent browser (multiple profiles), an encrypted vault the model never reads (browser_login, TOTP 2FA), and its own MCP endpoint for Claude Code or any MCP client; created in 1-3 s; live view of desktop, window or browser with takeover, embeddable in your app; REST API and platform MCP for one machine per customer with machine-scoped tokens and per-customer usage; warm pools, schedules, event webhooks; light or full (WebGL) browser; VPN exit countries; installs apt, .deb, AppImage, Flatpak |
| Computer-use desktop | Computer-use desktop | Managed: $8 for the first month, then $15/month for one hosted computer (size notManaged: $8 for the first month, then $15/month for one hosted computer (size not stated), with $5 starter AI credits once; 7-day refund; self-hosting free | Flat monthly plan; self-hosted asleep computers use disk only | Computers sleep and wake; files, Chromium profile and saved logins stay on the computer'sComputers sleep and wake; files, Chromium profile and saved logins stay on the computer's own Docker volume across sleep, wake and restart; asleep computers use disk only (self-hosting docs; hosted sleep policy not stated) | Container | Aug 2026 | Persistent Linux desktop with Chromium, files and saved logins for AI agents; MCP server for Claude Code and CursorPersistent Linux desktop with Chromium, files and saved logins for AI agents; MCP server for Claude Code and Cursor; Drive web UI with built-in AI or your own OpenAI/Anthropic key; default computer 1 CPU / 2 GB (self-hosted); open source, control plane and image AGPL-3.0, MCP/CLI/web MIT, 21 GitHub stars; the team is building an 'AI cofounder' sales agent on top |
| Computer-use desktop | Computer-use desktop | Site could not be loaded; the Product Hunt listing describes pay-as-you-go with a freeSite could not be loaded; the Product Hunt listing describes pay-as-you-go with a free trial but gives no prices | — | — | — | 2025 | Unconfirmed, from the Product Hunt listing only: virtual Android phones in the cloud for running AI agents in bulkUnconfirmed, from the Product Hunt listing only: virtual Android phones in the cloud for running AI agents in bulk, several devices in parallel, controlled through an API. The input row also named two other products: Phones Cloud (Product Hunt only, no website listed, a cloud Android work phone for teams with no mention of agents) and CaIoT DeviceHub/DeviceLab (caiot.co, real Android devices in the cloud for app testing and QA, Canada; the caiot.co homepage does not mention AI agents; hub.caiot.co blocked by Cloudflare) |
| Computer-use desktop | Computer-use desktop | $0.004 per phone-minute ($0.24/h), never more than $0.80 per phone per day$0.004 per phone-minute ($0.24/h), never more than $0.80 per phone per day; minimum top-up $10; Free 30 minutes and 100 MB traffic; Developer $29/mo (5,000 minutes, 10 GB traffic); Startup $99/mo (25,000 minutes, 25 GB); Scale $349/mo (100,000 minutes, 100 GB); Enterprise custom; extra residential traffic packs; phone size not stated | $0 (stopped phones are free) | Phones stop and start; a stopped phone costs nothingPhones stop and start; a stopped phone costs nothing; each phone keeps its own storage and settings until it is deleted | Container | Oct 2026 | Cloud Android phones for agents and app testing in 45 countries, ready in about a minuteCloud Android phones for agents and app testing in 45 countries, ready in about a minute; built-in residential network traffic or your own proxy; live screen, screenshots and UI tree; hosted MCP, Python SDK and MCP server (Clousd-Android/clousd-sdk, MIT), agent skill; team roles |
| Computer-use desktop | Computer-use desktop | $0.05 | $0.01/hr while stopped or suspended | Machines run until stopped or terminated; stop/start keeps state (stopped or suspendedMachines run until stopped or terminated; stop/start keeps state (stopped or suspended bills storage only); optional auto-destroy TTL 5 min to 7 days from creation; when the wallet runs out the VM is auto-stopped, never destroyed; terminate is irreversible | VM | Jul 2026 | Linux and Windows VMs with optional desktop; actions (click, type, scroll), screenshots, browser opsLinux and Windows VMs with optional desktop; actions (click, type, scroll), screenshots, browser ops, terminal (bash or PowerShell), file ops; SSH key + VNC password via connection endpoint; Predict API returns mouse/keyboard actions from a screenshot; Agents/Runs API runs the computer-use loop for you, BYOK model or managed; schedules with cron and webhooks; MCP server @coasty/mcp; REST with Python, Node and Go examples; claims #1 on OSWorld (82%) |
| Computer-use desktop | Computer-use desktop | $0.04 | Billing for paused sessions not statedBilling for paused sessions not stated; paid plans include 20-120 GB of persistent storage | Ephemeral by default (free-tier sessions start clean, 60-minute limit)Ephemeral by default (free-tier sessions start clean, 60-minute limit); paid plans keep a desktop's storage between sessions; pause freezes a desktop and resume brings back the same windows; snapshots and clones; pinned setups kept pre-warmed; closing an ephemeral session destroys its encryption key | — | Mar 2026 | Multiplayer Linux cloud desktop in the browser (P2P streaming, a cursor and keyboard per person)Multiplayer Linux cloud desktop in the browser (P2P streaming, a cursor and keyboard per person); REST API (OpenAPI 3.0) to boot sessions, exec shell commands, capture the screen, drive mouse and keyboard, snapshot, pause and resume; MCP server at api.cyqle.in/mcp for Claude Code, Codex, Gemini CLI, Qwen Code and Cursor; root access; built-in agent Factor; human takeover mid-run; mobile touch client |
| Computer-use desktop | Computer-use desktop | Every model the same: Monthly $50 per phone for 30 days (paid once, no auto-renewEvery model the same: Monthly $50 per phone for 30 days (paid once, no auto-renew; includes 2 GB mobile traffic and $10 of agent use); Hourly $0.28 per phone-hour, billed per minute while running ($200/month if never stopped); extra traffic $2/GB, own proxy free; agent use past the allowance $0.50 per million input and $2 per million output tokens; prepaid balance by card or crypto; trial $7 credit plus $1 agent use for 7 days. Not sized in vCPU/GiB | $0 while an hourly phone is stopped (apps and data kept); monthly phones prepaid | Hourly phones: stop the phone and billing stops, apps and data are keptHourly phones: stop the phone and billing stops, apps and data are kept; monthly phones run 24/7 for a paid 30 days, then stop and keep apps and data until renewed; if the balance runs out, running phones stop, and a phone stopped for a week for lack of balance is deleted with its data | — | Oct 2026 | Anti-detect cloud Android phones (Android 13 to 16) marketed for running many TikTok, InstagramAnti-detect cloud Android phones (Android 13 to 16) marketed for running many TikTok, Instagram, Facebook and X accounts, plus app testing and ad verification; passes Play Integrity basic, device and strong checks; device identity re-checked every 6 hours; US mobile IPs in Phoenix, Los Angeles, Charlotte, Atlanta, Chicago and Philadelphia; any Android app from Play Store or APK; MCP server (OAuth) for Claude, ChatGPT, Claude Code and Codex, REST API, TypeScript SDK, built-in agent runs over up to 50 phones; browser console |
| Computer-use desktop | Computer-use desktop | No price: billing is not enabled; compute is measured and shown as "not billed"No price: billing is not enabled; compute is measured and shown as "not billed"; spend caps apply to the measured cost; invited teams only | Not billed | A desktop closes when you close it, at its lifetime (default 3,600 sA desktop closes when you close it, at its lifetime (default 3,600 s, max 14,400 s) or after 1,800 s idle (settable idle_timeout); env.save() keeps a saved state (idle lease 1,800 s by default, max 5 per host, about 0.6 GB each) and env.fork(n) makes up to 8 copies on the host; reset restores the checkpoint | — | Oct 2026 | Isolated OSWorld Ubuntu desktops (1920x1080) for computer-use RL rollouts and evaluationsIsolated OSWorld Ubuntu desktops (1920x1080) for computer-use RL rollouts and evaluations; many desktops share one running OS and application set; fork N copies of one state for GRPO-style group rollouts; task verifiers with a per-episode gate report on restore and isolation validity; Python SDK (drums-sdk), HTTP API, hosted MCP server (34 tools) for Claude Code, Cursor and Codex; prime-rl and verifiers adapters; screenshot viewer (observation only); pilot host limit 10 desktops; drumsd host daemon open source (Apache-2.0) with a CRIU fork |
| Computer-use desktop | Computer-use desktop | Usage-based on simulator minutes: $0.10 per minute for iOS and Android ($6.00/h)Usage-based on simulator minutes: $0.10 per minute for iOS and Android ($6.00/h); the Starter, Production and Enterprise plans include 60 minutes a month until 2027; available on paid plans; sessions use the account's job concurrency and queue when it is full. Not sized in vCPU/GiB | Billed per simulator minute until the session stops (idle timeout optional)Billed per simulator minute until the session stops (idle timeout optional); $0 after it stops | A session runs until stopped or until its maximum duration (115 min on paid plansA session runs until stopped or until its maximum duration (115 min on paid plans, 40 min on Free; the docs' hard cap is that plus 5 min) or an optional idle timeout; when it stops, the simulator and controller endpoints are gone, and the recording, screenshots and logs are kept | VM | Nov 2026 | Cloud iOS Simulators (iPhone, iPad) and Android Emulators (phones, foldablesCloud iOS Simulators (iPhone, iPad) and Android Emulators (phones, foldables, tablets) that coding agents start with one command (eas simulator) or the REST API, install a build on, drive with agent-device, Argent or Appium, and leave proof (screenshots, session recording, logs) on PRs; live browser view with human takeover; no Mac or nested virtualization needed, works from Linux, WSL or a cloud sandbox; any app that runs on a simulator (React Native, Swift, Kotlin, Flutter, Unity); subset of Maestro flows; eas-simulator skill for Claude Code, Codex and Cursor; GitHub Actions and Expo Workflows |
| Computer-use desktop | Computer-use desktop | $0.1656 | Boxes are removed at expiry; no paused state described | A box lives for expiresIn (default 60 min, settable) and is then removedA box lives for expiresIn (default 60 min, settable) and is then removed; keepAlive (default 0) extends the expiry whenever the box is used; a device hosts one box at a time and you terminate it before starting another; VM Linux boxes can start from a named snapshot | Container | Mar 2025 | Environments for agents to operate Android phones, a Chrome browser (Playwright, multi-tab)Environments for agents to operate Android phones, a Chrome browser (Playwright, multi-tab), Linux desktops and code sandboxes; cloud physical Android devices or register your own; live view; UI actions; app install; run code in several languages; MCP and CLI (GitHub babelcloud/gbox, Apache-2.0, 183 stars) for Claude Code, Cursor and VSCode; OSWorld provider; homepage now promotes RLG-ALE ("Agent's Last Exam") GUI-agent gyms and datasets with on-premise, air-gapped deployment |
| Computer-use desktop | Computer-use desktop | Cloud phone use $0.007/min by default ($0.42/h)Cloud phone use $0.007/min by default ($0.42/h), capped at $1.20 per device per day on paid subscriptions; monthly rental $29.9 per device per month (unlimited minutes); parallels $39.9 per parallel per month (run that many phones without per-minute charges); time add-ons from 2,000 minutes that never expire; plans Free (2 profiles, 1 seat), Base and Pro (their prices are filled in by script and not readable in the page); Black Friday discounts until Nov 30. Not sized in vCPU/GiB | Billing is per minute of use; profiles count toward the plan's profile limit | Billed per minute of use unless covered by a monthly rental or a 'parallel' slotBilled per minute of use unless covered by a monthly rental or a 'parallel' slot; phones are created, started, stopped and deleted through the API, and profiles persist | — | Apr 2026 | Antidetect cloud phones and multi-account browser for social-media marketing (TikTok, Instagram, YouTube, RedditAntidetect cloud phones and multi-account browser for social-media marketing (TikTok, Instagram, YouTube, Reddit, X...), with RPA automation, a synchronizer, ADB, proxy management and an API; official GeeLark CLI (repo created 2026-05-20, 36 stars) and an experimental 'Awesome GeeLark Skill' that lets AI agents (OpenClaw, Claude, GPT) create, start and stop phones, install apps and run RPA tasks on 9 social platforms (Apache-2.0, 7 stars); GeeLark AI image and video tools |
| Computer-use desktop | Computer-use desktop | iPhone SE (3rd generation) $60/month; Mac mini M4 (16 GB RAM, 256 GB storage) $90/monthiPhone SE (3rd generation) $60/month; Mac mini M4 (16 GB RAM, 256 GB storage) $90/month; plus tax; each device its own monthly subscription, cancel anytime. No 2 vCPU/4 GiB size | Same as running (flat monthly rental) | Rented month to month; the device keeps appsRented month to month; the device keeps apps, files and sign-ins while the rental is active; on release (or 72 h after a failed renewal) the data is erased and the device and controller reset to factory settings; no hourly or daily rental | VM | Nov 2026 | Real iPhones and Macs for computer-use agents (Claude Code, Codex, Cursor, OpenClawReal iPhones and Macs for computer-use agents (Claude Code, Codex, Cursor, OpenClaw, Hermes or any agent) through an HTTPS API with OpenAPI spec, no MCP server needed; agent signs in by OAuth device flow and the human approves and pays in Stripe; live JPEG screen stream plus tap, type, swipe and click; iPhones can add an eSIM plan with SMS for verification texts and iMessage; install App Store apps and sign in to iCloud; build in Xcode on the Mac; human can take over from the browser; up to eight viewers, one controller; team sharing of rentals; no shell, filesystem mount or clipboard API |
| Computer-use desktop | Computer-use desktop | The CLI's onboarding claims a 'trial cloud phone'; no prices published | — | — | — | Jun 2026 | MIT-licensed CLI and MCP server for cloud Android phones: claim a trial phone, live relay viewerMIT-licensed CLI and MCP server for cloud Android phones: claim a trial phone, live relay viewer, a snapshot-then-act loop (accessibility tree, tap, type, keys), file push, and a scaffolded agent workspace with MCP config and skills; also drives local adb devices and emulators; macOS, Linux, Windows; repo has 0 stars and one day of commits (2026-06-05); the npm package 'handheld' is a reserved 0.0.0 placeholder whose homepage, mobileuse.dev, is also a dead Vercel deployment |
| Computer-use desktop | Computer-use desktop | Mac App $299/year; Helix for Linux $199/yearMac App $299/year; Helix for Linux $199/year; Helix Cloud $499 per user per month (30 concurrent agent desktops, 5,000 tasks a month, 10 projects; 72-hour trial with card); Enterprise pilot on your Kubernetes from $75K; Sovereign Server $175K one-time; no per-desktop hourly price | — | The desktop stays running so another person can pick up the agent's work where it left off | — | Oct 2023 | Each agent gets a GPU-accelerated streaming Linux desktop with browser, terminal, files and GUI appsEach agent gets a GPU-accelerated streaming Linux desktop with browser, terminal, files and GUI apps; 15 or more isolated desktops per node; ephemeral branch-scoped git keys; bring Claude Code, Codex, Gemini or open models; Helix Cloud, Kubernetes and Linux (Helm, air-gap ready), Mac app, or the Sovereign Server (8x RTX PRO 6000); SOC 2 Type II and ISO 27001; GitHub helixml/helix, 831 stars |
| Computer-use desktop | Computer-use desktop | Free $0 (5 environment-hours/month, 30-minute runtime, 1 concurrent)Free $0 (5 environment-hours/month, 30-minute runtime, 1 concurrent); Developer $29/month (100 environment-hours, no runtime cap, 5 concurrent); Enterprise custom (unlimited hours and concurrency, custom application environments) | $0 compute while stopped (billing pauses); storage cost not stated | Stop archives the sandbox: files persist and billing pausesStop archives the sandbox: files persist and billing pauses; resume restores it on fresh hardware; destroy permanently deletes a stopped environment; Free plan caps each environment at 30 minutes | Browser | Sep 2026 | Managed desktop application environments for computer-use agents, currently KiCad and FreeCADManaged desktop application environments for computer-use agents, currently KiCad and FreeCAD; health monitoring of the application itself; artifact checkpoints with recover(); screenshots, mouse, keyboard and file transfer (4 MiB per file); persistent Python cells on the desktop (64 KiB, 120 s each); KiCad live PCB API via kipy; session recordings up to 512 MiB or 6 h; Python SDK isle-sdk and REST API; GPT-6 Astra example; any model or agent framework; open-source arc-cua (MIT, 326 GitHub stars), a macOS action layer for computer-use agents |
| Computer-use desktop | Computer-use desktop | Community Edition free, up to 5 concurrent sessions, for individuals, non-profitsCommunity Edition free, up to 5 concurrent sessions, for individuals, non-profits, non-commercial use and evaluation; Basic, Essentials and Premier editions by contact (unlimited sessions), self-hosted or cloud-hosted by Kasm; no public prices | Self-hosted: your own servers; a paused session keeps using memory | Admin-set per group: keepalive_expiration (seconds without a client keepalive) thenAdmin-set per group: keepalive_expiration (seconds without a client keepalive) then delete, stop (keeps disk) or pause (keeps disk, processes and memory, still using RAM; container sessions only); session_time_limit expires a session; idle_disconnect drops the connection after N idle minutes; optional persistent profiles (roaming home directory) | Container | Aug 2021 | Streams containerized Linux desktops, browsers (Chrome, Firefox, BraveStreams containerized Linux desktops, browsers (Chrome, Firefox, Brave, Edge and others) and apps to a web browser through KasmVNC (GPL-2.0, 5,388 GitHub stars); open-source workspace images; browser isolation and web filtering; session recording, sharing, DLP; GPU passthrough; autoscaling Docker agents and VMs; single-server, multi-server and Kubernetes deployment, multi-region zones; Developer REST API; Windows desktops via RDP; markets "Secure Agentic AI" environments for computer-use agents but names no agent API or MCP |
| Computer-use desktop | Computer-use desktop | Credits: $0.04 each within the Entry and Individual monthly allowance ($0.032 annualCredits: $0.04 each within the Entry and Individual monthly allowance ($0.032 annual; overage $0.05; Trial overage $0.06). iOS Simulator 2 credits/min ($0.08/min); Android Emulator 1 credit/min ($0.04/min); Xcode 2/min active ($0.08) and 0.4/min warm ($0.016); Android Build 1/min active ($0.04) and 0.2/min warm ($0.008). Plans: Trial $0 (120 one-time credits), Entry $20/mo (500 credits), Individual $200/mo (5,000 credits), Company custom; annual saves 20%. Not sized in vCPU/GiB | Warm build instances $0.016/min (Xcode) or $0.008/min (Android build)Warm build instances $0.016/min (Xcode) or $0.008/min (Android build); deleted instances $0 | Instances live until deleted or a timeout fires: an inactivity timeout (simulatorsInstances live until deleted or a timeout fires: an inactivity timeout (simulators, emulators and Xcode default to the organization's setting, Gradle to 10 minutes; overridable, e.g. 10m or 3h) and an optional hard timeout (default none); Xcode and Android build instances bill a lower 'warm' rate between builds; disk snapshots restore build state across instances; labels let a create call reuse a warm instance | — | Sep 2025 | Remote Xcode builds (source synced on build, S3-backed cache, signing and App Store Connect upload)Remote Xcode builds (source synced on build, S3-backed cache, signing and App Store Connect upload), iOS simulators (iPhone, iPad, Apple Watch) and Android emulators streamed to a browser; accessibility tree, 60 fps video, log streaming, batch actions; Appium, Maestro, Detox and Bazel remote execution; per-instance and remote MCP servers; guides for Claude Code on the web and for agents in untrusted sandboxes; PR preview links; jurisdictions us, eu, as; domain blocklist, egress IPs, private tunnels; TypeScript, Go, Java and Python SDKs; customers include Replit; unlimited concurrency subject to abuse quotas |
| Computer-use desktop | Computer-use desktop | Plans with included instance hours: Free $0 (5 instance hours, 5 agent creditsPlans with included instance hours: Free $0 (5 instance hours, 5 agent credits, 2 concurrent instances); Basic $19/month (50 instance hours, 25 agent credits, 5 concurrent); Pro $49/month (250 instance hours, 50 agent credits, 10 concurrent); Enterprise custom; no overage rate on the homepage | — | Sandbox lifetime is set at creation: default 1 hourSandbox lifetime is set at creation: default 1 hour, extendable up to 1 day (Python SDK README); restart and delete; agent tasks can be paused, edited and resumed | — | Jul 2025 | Hosted computers and phones for GUI agents: Windows (shape beijing-2c-4g-cpu)Hosted computers and phones for GUI agents: Windows (shape beijing-2c-4g-cpu), Linux (beijing-2c-4g-cpu-linux) and Android (acep-shenzhen-enhanced, acep-wenzhou-common-pro) per its official skills repo; live desktop stream with human takeover and recordings; MCP server; SDKs for Python, TypeScript, Go and Rust (MIT) plus a Dify tool; mouse, keyboard, touch and app actions, screenshots, process execution, file transfer (HTTP, S3, MinIO) and HTTP port mapping; create a sandbox from an existing one's image; open-source Lybic agent framework (Apache-2.0, 119 stars); operated by TINGYU (SINGAPORE) TECH PTE. LTD.; docs and dashboard on lybic.cn |
| Computer-use desktop | Computer-use desktop | Monthly resource pools, paid plan required: Solo $29/mo (up to 4 computersMonthly resource pools, paid plan required: Solo $29/mo (up to 4 computers, pool 4 vCPU / 8 GB / 80 GB, max 4 vCPU per computer); Studio $99/mo (16 computers, 16 vCPU / 32 GB / 320 GB); Swarm $399/mo (32 computers, 64 vCPU / 128 GB / 640 GB); built-in agent credits $5/$10/$50 per month when enabled, billed at 1.3x Anthropic list price | Included in the monthly plan; a stopped or suspended computer frees pool vCPUIncluded in the monthly plan; a stopped or suspended computer frees pool vCPU and memory but keeps its disk in the storage pool | Persistent: files, logins and installed software stayPersistent: files, logins and installed software stay; stop or suspend a computer (suspend saves memory to disk) to free its vCPU, memory and running slot, while its disk still counts against the plan's storage pool; snapshots on demand or a daily schedule (7 days to 12 months of history by plan); ephemeral computers from the SDK clean themselves up; built-in agent runs stop after 60 min | VM | Jul 2026 | Persistent Linux desktops (Debian Xfce, Omarchy/Hyprland) for computer-use agentsPersistent Linux desktops (Debian Xfce, Omarchy/Hyprland) for computer-use agents; templates with Claude Code, Hermes Agent, OpenClaw, ten coding-agent CLIs, or the Claude Managed Agents self-hosted worker preinstalled; custom templates; screenshot, click, type, key and exec in the graphical session; live view embeddable with one tag (view-only or interactive); public HTTPS URLs under mandala.dev; secrets and a credential gateway; REST API, Python and TypeScript SDKs (MIT), hosted and local MCP server, CLI; workspaces with scoped API keys; resize and clone |
| Computer-use desktop | Computer-use desktop | Hobby $0 (10 device minutes a day, 1 concurrent session)Hobby $0 (10 device minutes a day, 1 concurrent session); Starter $29/mo (200 minutes, 2 sessions, overage $0.15/min up to your spend cap); Team $299/mo (2,500 minutes, 5 sessions, 10 members, overage $0.15/min); Enterprise custom (dedicated device pools); work-email sign-ups start with 100 free minutes. Not sized in vCPU/GiB | N/a (billed only from allocation to release) | Billed by the second from device allocation to releaseBilled by the second from device allocation to release, with a 5-minute minimum per session; no idle timeout stated | — | Mar 2025 | Mobile MCP (Apache-2.0, 10,571 stars), mobilecli (FSL) and Mobilewright ('Playwright for mobile'Mobile MCP (Apache-2.0, 10,571 stars), mobilecli (FSL) and Mobilewright ('Playwright for mobile', Apache-2.0) let Claude Code, Cursor, Copilot, Gemini and other MCP agents drive real devices, simulators and emulators; cloud sessions over REST API, MCP, CLI or dashboard; upload your .ipa or .apk; session traces and screen recordings; the open-source tools are free on your own devices |
| Computer-use desktop | Computer-use desktop | Cloud Android 14 phone $0.01 per phone-minute ($0.60/h)Cloud Android 14 phone $0.01 per phone-minute ($0.60/h), billed per second while ready with a 1-minute minimum, boot time free; real Android device $0.01/min; iPhone and iPad $0.05/min; dedicated devices from $79/month; residential IP $8/GB, mobile carrier IP $15/GB; free credit for new workspaces; no plans or seats. Not sized in vCPU/GiB | $0 (parked phones are free) | A phone parks after 5 idle minutes, and after 15 minutes of running unless max_durationA phone parks after 5 idle minutes, and after 15 minutes of running unless max_duration asks for up to 3 hours; a parked phone keeps its installed apps, data and logins and costs nothing; start resumes the same phone | — | Nov 2026 | REST API and hosted MCP server (15 tools: create, start, park and delete phones, install apps, observe, actREST API and hosted MCP server (15 tools: create, start, park and delete phones, install apps, observe, act, screenshot...); screens read as numbered elements, up to 20 actions per call; ready in 19.4 s in its November 2026 benchmark; live view; country, locale, timezone and GPS settings; aimed at cloud agents without KVM (Claude Code on the web, Codex, Cursor); integrations for OpenAI Agents, LangGraph, CrewAI and browser-use; OpenAPI spec; GitHub fifteenA/mobilebox (MIT, 1 star) |
| Computer-use desktop | Computer-use desktop | Free $0 (60 device minutes a month, 500 credits = $5 of agent usageFree $0 (60 device minutes a month, 500 credits = $5 of agent usage, 1 GB residential traffic, 1 concurrent phone); Developer $20/mo (600 min, 5 phones); Startup $99/mo (3,000 min, 25 phones); Scaleup $299/mo (9,000 min, 100 phones); extra device minutes $0.03 ($1.80/h), credits $0.01, residential traffic $10/GB; dedicated Android $50/mo, Android + dedicated IP $64/mo yearly ($80 monthly), iPhone + dedicated IP $160/mo yearly ($200 monthly), your own device $4/mo yearly ($5 monthly); Enterprise custom. Not sized in vCPU/GiB | $0 device minutes while stopped (apps and data kept)$0 device minutes while stopped (apps and data kept); dedicated phones are flat monthly | Plan phones use device minutes only while powered onPlan phones use device minutes only while powered on, and a stopped phone keeps its apps and data; dedicated phones stay on around the clock | — | Apr 2025 | Cloud phones for agents (Claude Code, Codex, Cursor, OpenClaw, Hermes or its own mobilerun VA) over REST, PythonCloud phones for agents (Claude Code, Codex, Cursor, OpenClaw, Hermes or its own mobilerun VA) over REST, Python, TypeScript and MCP; receive-only phone numbers and inboxes for SMS and email verification; residential network in a chosen country and city; live view with screenshots; webhooks; connect your own iPhone over USB from Chrome on a Mac; open-source mobilerun framework (MIT, GitHub droidrun/mobilerun 10,043 stars); claims 98.34% on iOSWorld and 100% on AndroidWorld |
| Computer-use desktop | Computer-use desktop | Prices not published (the List prices API needs an account)Prices not published (the List prices API needs an account). Paid from a prepaid balance with a hold at create time; VMs are charged per reserved Mac by the hour, and a second VM in the Mac's other slot is included; bare metal is charged per Mac in 24-hour windows; access by request | Same as running (charged per reserved Mac while it is held) | No pause or suspend documented: a VM runs until you delete it (delete removes the VM andNo pause or suspend documented: a VM runs until you delete it (delete removes the VM and its disk); a bare-metal Mac is leased until released and is wiped before anyone else uses it | VM | Nov 2026 | Elastic Apple hardware for agents: computer-use RL data, persistent macOS environments for virtual coworkersElastic Apple hardware for agents: computer-use RL data, persistent macOS environments for virtual coworkers, Apple-platform app testing, open-model inference on Apple silicon; machine classes such as m4-24g (blog mentions 512 GB M5 Ultra clusters); bring your own VM images; admin and sudo in VMs; SSH or virtual desktop (Screen Sharing) via the mthr CLI; portal, REST API with OpenAPI, Kubernetes provisioning; event webhooks; signed machine identity tokens; customer-managed Macs at Mount Thor sites; Tier 1 colocation in US metros, Europe and Asia coming soon |
| Computer-use desktop | Computer-use desktop | $0.90/h per Mac, billed per minute while the Mac is ready, one-minute minimum$0.90/h per Mac, billed per minute while the Mac is ready, one-minute minimum; prepaid credit in $10 packs, no subscription, never expires; bonus credit on bigger purchases (+5% from $20 to +50% at $1,000; as low as $0.60/h); NoMac Simulator also $0.90/h; a session Mac is 6 cores and 8 GB, so no 2 vCPU/4 GiB size | $0 when stopped; the 20 GB drive is included | A Mac stops after 10 minutes idle, at its time limit (2 h by default, 8 h at most)A Mac stops after 10 minutes idle, at its time limit (2 h by default, 8 h at most), at its spend cap ($5 by default) or when credit runs out; stopping deletes the VM and everything outside ~/drive; an encrypted 20 GB drive per Mac persists for the next session | VM | Jul 2026 | Real macOS with Xcode 27, Homebrew, Node, Python and CocoaPods preinstalledReal macOS with Xcode 27, Homebrew, Node, Python and CocoaPods preinstalled; passwordless sudo and auto-answered permission dialogs; iOS Simulator with screenshots and recordings; signing, TestFlight uploads and App Review submission; MCP, CLI, REST API and SSH; desktop in the browser; secrets; per-user Macs through the macOS Platform SDK (TypeScript, Python) with webhooks; NoMac CI runner for GitHub Actions; built by Gary Miklos |
| Computer-use desktop | Computer-use desktop | By the hour from credit, per second: Small 0.5 vCPU/4 GB/20 GB $0.023/h (API only)By the hour from credit, per second: Small 0.5 vCPU/4 GB/20 GB $0.023/h (API only); Medium 2 vCPU/8 GB/40 GB $0.055/h; Large 4/16/80 $0.11/h; XL 6/32/150 $0.202/h; 2XL 8/64/300 $0.368/h (on request); plans Starter $29/mo (1 Medium always on), Growth $99/mo (4 Mediums), Scale $399/mo (16 Mediums), yearly 20% off; add-on Medium $32/mo; Windows +$15/mo license; Mac from $100/mo; physical iPhone 12 $65/mo, iPhone 17 $129/mo; GPU computers (AMD V710 or Intel B50) $47-$240/mo, hourly at month x 1.35 / 730; residential IP data $3/GB past plan; $3 starting credit for new accounts; no 2 vCPU/4 GB size | Stopped hourly computer: disk only, $1 per 20 GB-month (Medium $0.0027/h)Stopped hourly computer: disk only, $1 per 20 GB-month (Medium $0.0027/h); suspended computers on a plan are covered by the plan | Never auto-suspended; runs until you stop or delete it, no time limitNever auto-suspended; runs until you stop or delete it, no time limit; on a plan, suspend saves memory and resume restores open windows; hourly computers stop instead (disk kept, billed for disk only); optional end time 1-720 h then deleted; out of credit: computer stops, files kept 7 days then deleted | VM | Mar 2025 | Linux, Windows and macOS desktops plus physical iPhones through one API; boots in 500 ms claim; SSH with your own keyLinux, Windows and macOS desktops plus physical iPhones through one API; boots in 500 ms claim; SSH with your own key; live desktop view and embeddable screen; suspend with memory; live resize; clone and fork while running; daily backups (Linux, Windows); templates; public URLs and custom domains; residential IPs for Chrome; secrets and scoped API keys; GPU computers where enabled; servers in the US and London; Python and TypeScript SDKs, CLI, MCP; agents can sign up for their own account by API |
| Computer-use desktop | Computer-use desktop | Basic: cloud Android emulators 5¢ per phone-minute ($3.00/h), $5 starting creditBasic: cloud Android emulators 5¢ per phone-minute ($3.00/h), $5 starting credit, top-ups $5-$200; Plus $20/month with 1,000 phone-minutes, then 4¢/min; Cloud iPhone $80/month ($100 with eSIM), annual $64/month ($80 with eSIM); Scale (dedicated real iPhones or Androids) by contact. Not sized in vCPU/GiB | Android: $0 after the session ends (only the saved phone persistsAndroid: $0 after the session ends (only the saved phone persists, no stated storage charge); Cloud iPhone: flat monthly | Each session has a fixed timeout_seconds deadline (default 300 sEach session has a fixed timeout_seconds deadline (default 300 s, no unlimited mode or extension) that includes provisioning; billing runs from ready until closure; a temporary Android phone keeps nothing; one saved Android phone per account keeps apps, logins and files between sessions; a Cloud iPhone is never wiped between sessions, only on reset | — | Aug 2026 | Open-source (MIT) harness that lets an agent tap, type and scroll phone appsOpen-source (MIT) harness that lets an agent tap, type and scroll phone apps, reading the screen by vision OCR or the accessibility tree; Phone Harness Cloud gives hosted Android emulators and real iPhones through an API, MCP server, live viewer and adb connect; real iPhones get a real number with eSIM for SMS and 2FA; 8 phones per account at once; Android boot about 80 s; OpenClaw guide; GitHub ShawnPana/phone-harness 4,990 stars |
| Computer-use desktop | Computer-use desktop | $0.06 per phone-minute ($3.60/h), billed per second with a one-minute minimum$0.06 per phone-minute ($3.60/h), billed per second with a one-minute minimum; Starter $20/mo (includes $20 of phone time, 5 phones at once) after a 7-day trial with card; Pro $99/mo (includes $99, 20 phones at once); Enterprise custom; phone size not stated | $0 (parked phones cost nothing) | A phone parks after inactivity (5 minutes by default per the pricing FAQA phone parks after inactivity (5 minutes by default per the pricing FAQ, 10 per the homepage; at most 60), at its session limit (15 minutes by default, up to 3 h) or on request; a parked phone keeps its apps, files and signed-in accounts | Container | Oct 2026 | Cloud Android phones for agents: the screen read as a numbered list of elements, tap by label, type, swipe, open appsCloud Android phones for agents: the screen read as a numbered list of elements, tap by label, type, swipe, open apps; install APKs, move files, set GPS, locale and timezone; 48 countries; recordings; REST API, TypeScript SDK, CLI and hosted MCP; every charge tied to a session and a monthly spending limit; no phone number or SMS; made by Steelworks Software LLC |
| Computer-use desktop | Computer-use desktop | Cloud phone monthly $20, $24, $28 or $32 by model (unlimited minutes)Cloud phone monthly $20, $24, $28 or $32 by model (unlimited minutes); cloud phone with Real phone identity $39/month; on-demand cloud phone $0.024/min ($1.44/h) from prepaid balance; Robot hand + iPhone $149/month per robot, prepared within 7 days; Enterprise by call. Not sized in vCPU/GiB | Per-minute phones: same as running until returned; monthly phones: flat price | Per-minute cloud phones bill from ready until you return them or the prepaid balance runsPer-minute cloud phones bill from ready until you return them or the prepaid balance runs out (10 minutes of rate held at start); monthly cloud phones are always on and keep apps and logins between runs; device leases expire after 10 minutes unless renewed | — | Sep 2026 | Phones for AI agents through an MCP server (12 tools: acquire, screen read, tap, type, swipe, open app, read SMS...)Phones for AI agents through an MCP server (12 tools: acquire, screen read, tap, type, swipe, open app, read SMS...), HTTP API and npm CLI; works with Claude, Cursor, Codex, Windsurf; Android cloud phones boot in about 4 s and scale to hundreds; regions chosen per cloud phone by country; real phones have carrier SIMs for SMS/2FA and pass device-integrity checks (no jailbreak or helper app); snapshots for clean eval phones; homepage counted 312 phones online (288 cloud, 24 real); docs say the robot-hand iPhone is bought in the console and is not yet in the API, MCP or CLI |
| Computer-use desktop | Computer-use desktop | Free $0 (30 minutes a month, 1 device); Hobby $2/month (200 minutes, 1 device)Free $0 (30 minutes a month, 1 device); Hobby $2/month (200 minutes, 1 device); Pro $10/month (unlimited minutes, 2 devices); Enterprise custom; no per-minute rate | N/a (sessions end; no paused state) | Sessions end after 10 minutes without device inputSessions end after 10 minutes without device input; a fresh device every session, with app data, logins and files discarded when it ends; share links last 3 days by default, up to 7 | VM | Sep 2026 | Cloud Android emulators for coding agents over a remote MCP server (api.ramus.dev/api/mcp) or ramus-cliCloud Android emulators for coding agents over a remote MCP server (api.ramus.dev/api/mcp) or ramus-cli; Claude Code, Codex, Cursor and ChatGPT; builds Expo, React Native, Flutter and Kotlin/Gradle apps from GitHub; UI tree, taps, logcat; share links and GitHub PR previews; no Play Store; iOS emulators 'coming soon' |
| Computer-use desktop | Computer-use desktop | Free $0/month (free monthly usage, 2 concurrent devices, 4 with pay-as-you-go)Free $0/month (free monthly usage, 2 concurrent devices, 4 with pay-as-you-go); Starter $250/month ($200/month annual) with $250/month of compute credits and 10 concurrent devices; Enterprise custom. Overage $0.15/min for iOS simulators ($9.00/h) and $0.12/min for Android emulators ($7.20/h); Mac computers $0.05/min ($3.00/h, on paid plans or on request). No vCPU/GiB sizes published | Billed per minute while a session or Mac runs (idle timeout optional)Billed per minute while a session or Mac runs (idle timeout optional); $0 after it stops, nothing kept | Device sessions take an idle timeout of 5, 15Device sessions take an idle timeout of 5, 15, 30 or 60 minutes and are flagged after 2 minutes idle; a Mac computer runs at most two hours from creation and is deleted when it stops, so nothing persists; no pause | Browser | 2024 | Runtime and verification platform for cloud coding agents: agents buildRuntime and verification platform for cloud coding agents: agents build, run and verify every change on cloud iOS simulators, Android emulators, web browsers and full macOS desktops streamed to the browser; Atlas keeps every session as a queryable record and app map; proof-of-changes reports on pull requests; CLI, REST API, Verify SDK for TypeScript and Python, Agent Skill; Appium driver; GitHub Actions jobs on single-use Macs (runs-on: revyl-macos); exec commands up to an hour; Capy uses Revyl for its cloud phones; SOC 2 Type II, SSO, RBAC on Enterprise |
| Computer-use desktop | Computer-use desktop | Free $0 (10 min device time/day, 1 device, shared pool, MCP)Free $0 (10 min device time/day, 1 device, shared pool, MCP); Manual $25/device/month (live control + MCP, unlimited device hours); Automation $125/device/month (adds Selenium/Appium grid, parallel runs, CI/CD); Web $50/slot/month (desktop browsers only); Enterprise from $125/device + $20/seat/month, 10-device and 10-seat minimum, billed annually; $5/month AI credit included; no vCPU/GiB sizes | Flat monthly per-device price; same whether used or not | Devices are rented per slot per month with unlimited device hours on paid plansDevices are rented per slot per month with unlimited device hours on paid plans; devices are handed back muted on release; Free tier gives 10 minutes of device time a day | Browser | May 2025 | Real-device cloud for QA teams and AI agents; MCP server (npm @robotactions/mcp, SSE endpoint) lets ClaudeReal-device cloud for QA teams and AI agents; MCP server (npm @robotactions/mcp, SSE endpoint) lets Claude, Cursor or any MCP agent drive real iOS/Android devices with hundreds of tools (gestures, location, network capture and mocking, logs, crash reports, VoiceOver/TalkBack); open-source Agent Skills; built-in AI test agent and chat agent working from Jira, TestRail, Azure DevOps; W3C WebDriver endpoint for Selenium, Appium, Playwright, WebdriverIO; video recording; any device model on request; 30-day free trial; GitHub repo krishtoautomate/robotactions holds only issue intake and release notes |
| Computer-use desktop | Computer-use desktop | iOS simulator $0.02 per active minute ($1.20/hr)iOS simulator $0.02 per active minute ($1.20/hr); Android emulator $0.02 per active minute; Xcode builds free during beta; $5 usage to start with no card, then first $15 of usage free each month with a card (up to 750 simulator minutes); no Linux sandbox or CPU/RAM pricing on the site now | $0 (metering stops when a session is released or times out) | Sessions are metered from create until released or until their inactivity timeout endsSessions are metered from create until released or until their inactivity timeout ends them (set per session, e.g. inactivityTimeout '60s'); idle sessions time out instead of running up the meter | — | — | Browser-streamed iOS simulators and Android emulators for mobile agentsBrowser-streamed iOS simulators and Android emulators for mobile agents; create, open-url, install app builds, screenshot, screen recording, release over API; TypeScript SDK @run-cloud/sdk and CLI runcloud; MCP server at mcp.run.cloud/mcp for Claude, ChatGPT, Claude Code and Codex; embeddable iframe sessions; cloud Xcode archive, sign and export to TestFlight |
| Computer-use desktop | Computer-use desktop | No prices published; early access; ruOS Lite browser desktop is free | — | Files and the AI's memory are saved automaticallyFiles and the AI's memory are saved automatically, and the desktop picks up where you left off, even on another device | — | — | Agentic Linux desktop opened in any browser, with Claude Code, Codex, VS Code and the ruvnet tools (ruflo, ruvectorAgentic Linux desktop opened in any browser, with Claude Code, Codex, VS Code and the ruvnet tools (ruflo, ruvector, ruview) preinstalled; MCP server ruos-computeruse-mcp (screenshot, mouse and keyboard via xdotool, shell, resolution); hosted MCP endpoint ruos.cognitum.one/mcp; ChatGPT plugin; part of the Cognitum platform |
ScrapybaraDiscontinued (site lists the computers-for-agents product as 2024-2025 history) Computer-use desktop | Computer-use desktop | No current pricing; the site describes the remote-desktop API as 2024-2025 history andNo current pricing; the site describes the remote-desktop API as 2024-2025 history and the old pricing URL (scrapybara.com/pricing) returns 404 | — | — | — | 2024 | Former product: remote desktops that AI agents can see and control through an API and SDKs (2024-2025)Former product: remote desktops that AI agents can see and control through an API and SDKs (2024-2025); company now builds Capy (capy.ai), an agent-swarm platform for engineering teams; docs.scrapybara.com redirects to docs.capy.ai; Python SDK repo last pushed 2025-09-08 |
| Computer-use desktop | Computer-use desktop | Active computer $0.005/min ($0.30/h); hibernated $0.0005/min ($0.03/h)Active computer $0.005/min ($0.30/h); hibernated $0.0005/min ($0.03/h); $0.003 per agent step (screenshot, click, type); $0.05 per file upload or download; Free $0/mo with $5 credit; Pro $20/mo with $20 credit; Enterprise custom, on your own VMs; VM size not stated, so no 2 vCPU/4 GiB price | $0.0005/min ($0.03/h) hibernated | Pause and resume, or stop and start, without losing progressPause and resume, or stop and start, without losing progress; hibernated computers billed at a lower rate; terminate deletes | VM | Apr 2025 | Hosted Windows computers for computer-use agents; Interactions API for screenshots, clicks, typing and keysHosted Windows computers for computer-use agents; Interactions API for screenshots, clicks, typing and keys; built-in Windows agent with human-in-the-loop; live streaming and takeover; install apps; file upload and download; Python and TypeScript SDKs usable as an MCP server; 'Ledra AI' does not appear on the site |
| Computer-use desktop | Computer-use desktop | Not published; waitlistPricing page | — | VMs boot from memory snapshots of pre-warmed golden imagesVMs boot from memory snapshots of pre-warmed golden images; checkpoint, restore, rollback and fork; no idle policy stated | VM | Jan 2026 | Desktop VMs for macOS, Windows, Linux, iOS and Android, plus microVMs; boots in seconds from memory snapshotsDesktop VMs for macOS, Windows, Linux, iOS and Android, plus microVMs; boots in seconds from memory snapshots; snapshots shared as URLs; SSH and a web VNC; built-in customizable agent plus Claude Code and Codex; computer-use tools (screen capture, shell, accessibility tree); GitHub App provisions a VM per PR; Linear, Slack, Discord, Raycast; API, CLI, VS Code extension, mobile app; SDKs for Python, TypeScript and Rust; custom golden images; self-hosted option for enterprise |
Symphony (symphon.co)Appears shut down (empty site; login host gone) Computer-use desktop | Computer-use desktop | Per the 2025-08-08 archived copy: Free $0 (30 GB storage, 4 GB memory)Per the 2025-08-08 archived copy: Free $0 (30 GB storage, 4 GB memory); Pro $19/month (permanent OS); Premium $29/month (API access for AI and screenshots, custom OS, 200 GB storage, 8 GB memory) | — | Per the 2025-08-08 archived copy: the free OS resets after 3 days of inactivityPer the 2025-08-08 archived copy: the free OS resets after 3 days of inactivity; paid plans keep a permanent OS and data | — | — | Per the archived copy: a desktop OS in the browser with an AI assistant that controls mouse and keyboardPer the archived copy: a desktop OS in the browser with an AI assistant that controls mouse and keyboard, runs terminal commands and opens apps (OnlyOffice, LibreOffice, VS Code, Discord, Slack); on 2026-11-14 symphon.co returned an empty page and login.symphon.co no longer resolved; 'meetOS1' from the lead could not be traced (meetos1.com does not resolve) |
| Computer-use desktop | Computer-use desktop | $1.14 | Billed on CPU and memory actually used while the computer runsBilled on CPU and memory actually used while the computer runs; it terminates after the inactivity timeout | Auto-terminates after an inactivity timeout (any action or keepalive resets it) and at aAuto-terminates after an inactivity timeout (any action or keepalive resets it) and at a maximum lifetime, both plan-dependent; persistent=True saves the environment's state on termination, restorable later by environment_id | Container | Apr 2026 | Lightcone computer-use API: cloud browser and Linux desktop computers driven by Tzafon's Northstar vision-languageLightcone computer-use API: cloud browser and Linux desktop computers driven by Tzafon's Northstar vision-language model (Tasks, OpenAI-compatible Responses and Chat Completions APIs) or directly through the Computers API; CDP access for Playwright; browser kind has residential proxies and stealth mode; human-in-the-loop login handoff; Python and TypeScript SDKs; CLI with MCP for coding agents; Northstar can also run fully on-device on a Mac (MLX 4-bit build, per the docs); earlier open-source browser-fleet tool Tzafon-WayPoint (2025-04-21) |
| Computer-use desktop | Computer-use desktop | List prices only from Volcengine sales ('contact the Volcengine sales team')List prices only from Volcengine sales ('contact the Volcengine sales team'); instances prepaid monthly or yearly, or pay as you go settled daily at the day's peak instance count × a daily unit price; bandwidth, cloud-disk storage and EIP billed separately. Sizes include 2 vCPU / 4 GB instances (samplePlus and samplePro, four per 8 vCPU/16 GB host), but no prices are published | Pay as you go bills each day's peak instance countPay as you go bills each day's peak instance count; whether a powered-off instance counts is not stated | Instances can be powered off and on, rebooted, resetInstances can be powered off and on, rebooted, reset, backed up and restored through the OpenAPI; pay-as-you-go resources can be started and stopped at any time; no idle auto-suspend stated | VM | Jun 2025 | Cloud Android phones (AOSP 10, 11, 12, 13 and 15) streamed to Android, iOS, Web/H5 and Windows clientsCloud Android phones (AOSP 10, 11, 12, 13 and 15) streamed to Android, iOS, Web/H5 and Windows clients; Mobile Use Agent (a GUI agent that operates mobile apps, with OpenAPI, CLI and MCP tool calls) and Computer Use Agent on cloud sandbox desktops; AI apps in the console run OpenClaw and Hermes against the phone (2026-06-11); local-storage hosts split 1 to 10 ways, or cloud-disk instances; ADB, proxy settings, custom images; sample-code repo volcengine/vePhone (27 stars, created 2022-10-19) |
| Computer-use desktop | Computer-use desktop | Pay-as-you-go $0.40 per Cloud PC hour (US geography)Pay-as-you-go $0.40 per Cloud PC hour (US geography), counting time spent running agent tasks in each hourly window, rounded up to the next full hour; always-available Cloud PCs add $5 per Cloud PC per month on top of usage; prices vary by geography; needs an Intune license, an Agent 365 license and a Windows 365 for Agents billing plan; Cloud PC size (vCPU, RAM) not stated in the agents docs | Pay-as-you-go: billed on time spent running agent tasksPay-as-you-go: billed on time spent running agent tasks; always-available: $5 per Cloud PC per month plus usage | Agents check a Cloud PC out of a pool and check it in when doneAgents check a Cloud PC out of a pool and check it in when done; on release it is reset and state is lost (docs say to checkpoint externally); a session with no MCP request for 30 minutes is evicted and reset; no maximum session length; the always-available add-on keeps Cloud PCs ready for instant checkout | VM | Jun 2026 | Windows Cloud PCs for computer-using agents, joined to Entra ID and managed in IntuneWindows Cloud PCs for computer-using agents, joined to Entra ID and managed in Intune; pooled check-out/check-in with 1 to 200 always-available Cloud PCs per provisioning policy; MCP servers for acquiring a Cloud PC and for actions (click, type, navigate, run commands) through an on-box CUA client; human live view and take-over in the same session; used by Copilot Studio computer use, Project Opal, Researcher computer use, Agent 365 and, per the architecture page, Manus; Defender, Purview, Global Secure Access and an Intune security baseline; gallery or custom images; provisioning takes 20-30 minutes; Graph APIs |
| Unified API | Unified API | No Agent Machines charge published: you bring provider and model keysNo Agent Machines charge published: you bring provider and model keys, and the providers bill you directly ('Agent Machines does not offer a unified bill'); the pricing page's $0.0648/vCPU-hour + $0.0072/GiB-hour are labelled illustrative, not a rate card | Billed by the chosen provider | Depends on the provider (FAQ): Daytona stop/start keeps files but restarts processesDepends on the provider (FAQ): Daytona stop/start keeps files but restarts processes; E2B pause/resume; Sprites suspends when idle; Vercel resumes from filesystem snapshots; the Worker's saved state persists across sleep and wake | — | May 2026 | Control plane for persistent agent 'Workers' that pairs an agent runtime (Claude Code, Codex, HermesControl plane for persistent agent 'Workers' that pairs an agent runtime (Claude Code, Codex, Hermes, OpenClaw) with a sandbox provider and a model path; worker presets (coding, code review, research, data, computer use, support); browser terminal with a live PTY; 161 SKILL.md entries and 39 MCP catalog entries; cron, logs, usage and artifacts; TypeScript SDK and hosted API; open source (MIT), 29 GitHub stars |
| Unified API | Unified API | Free, open source (MIT per the README and package.json; the repo has no LICENSE file)Free, open source (MIT per the README and package.json; the repo has no LICENSE file); sandbox providers bill separately | — | Set by the sandbox provider; the SDK provisions with findOrProvision and removes withSet by the sandbox provider; the SDK provisions with findOrProvision and removes with sandbox.delete | Container | Apr 2026 | TypeScript SDK that runs Claude Code, Codex or OpenCode inside a sandbox as each agent's own server process (CodexTypeScript SDK that runs Claude Code, Codex or OpenCode inside a sandbox as each agent's own server process (Codex app-server over JSON-RPC, OpenCode serve over HTTP/SSE, Claude Code SDK WebSocket) instead of non-interactive CLI calls, keeping approval flows and streaming events; normalized event stream; image presets built per provider (Docker tag, Modal image, E2B template, Daytona snapshot); skills, sub-agents, MCP servers, hooks, multimodal input; tracks Claude Code background tasks; v0.1.603; 166 GitHub stars; made by Twill, a cloud service for delegating coding-agent tasks |
| Unified API | Unified API | Free, open source (Apache-2.0); self-hosted with your own subscriptions and keysFree, open source (Apache-2.0); self-hosted with your own subscriptions and keys ('$0/session') | N/a (self-hosted; providers bill their own sandboxes) | Depends on the chosen sandbox; not described | Container | Apr 2026 | Self-hosted implementation of Anthropic's Managed Agents API (managed-agents-2026-04-01 spec)Self-hosted implementation of Anthropic's Managed Agents API (managed-agents-2026-04-01 spec), so the official Anthropic SDK works by changing baseURL; wraps 5 harnesses (Claude Code, Codex CLI, Gemini CLI, OpenCode, Factory droid) and 11 sandboxes; site lists cold starts from about 0.1 s (Daytona) to 3 s (Fly.io); web UI on localhost:4000; npx or Docker install; outcomes, multi-agent threads and memory 'coming soon'; same org as mvm (the mvm row); 0 GitHub stars, last push 2026-08-01 |
| Unified API | Unified API | Pay per run: the runner's listed price plus a 5% platform fee after your first top-up (0%Pay per run: the runner's listed price plus a 5% platform fee after your first top-up (0% while spending the $10 welcome credits); output normalization +$0.10/run; failed, blocked and parallel fallback attempts are billed; no seats or subscriptions; Enterprise custom | — | — | Browser | Aug 2026 | "OpenRouter for browser agents": one OpenAI-compatible Responses API (POST /api/v1/responses) over 50+ runners (hosted"OpenRouter for browser agents": one OpenAI-compatible Responses API (POST /api/v1/responses) over 50+ runners (hosted agents, browser infrastructure and open-source frameworks) in 250+ combinations; smart routing from a per-domain success scoreboard; on a block it races two or three runners in parallel; receipts with screenshots and per-attempt costs; drop-in wrappers for the Browserbase, Steel, Browser Use, Skyvern and Hyperbrowser SDKs (the Browserbase wrapper returns a proxied Browserbase CDP URL); MCP server, agent skill and agent self-signup; credential vault and MFA relay; 7-day run history |
| Unified API | Unified API | Free, open source (Apache-2.0); each provider bills its own sandboxesPricing page | Billed by the chosen provider | Depends on the provider; saved references recover sandboxes across serverlessDepends on the provider; saved references recover sandboxes across serverless invocations, and scoped cleanup can keep snapshots and volumes | — | Sep 2026 | Typed TypeScript SDK and CLI for creating, running code in and tearing down sandboxesTyped TypeScript SDK and CLI for creating, running code in and tearing down sandboxes; adapters for boxd, Freestyle, Daytona and E2B, plus experimental Boat, Runloop, Archil, exe.dev, Vercel and Modal; defineAdapter for other providers; CLI never prompts and has --json output for coding agents, plus an agent skill; OpenTelemetry traces (Sentry, Datadog); v0.15.0 on npm; 1 GitHub star |
| Unified API | Unified API | Free, open source (Apache-2.0) SDK and self-hosted control planeFree, open source (Apache-2.0) SDK and self-hosted control plane; you pay each backend provider at its own rates | N/a (library; each provider's own idle pricing applies) | The context manager destroys the sandbox on exitThe context manager destroys the sandbox on exit; the self-hosted control plane enforces maximum requested lifetimes and sandbox deadlines and cleans up after a client disconnects; snapshots on Docker, Tensorlake, Modal and boxd; reattach to a live sandbox from another process; otherwise the backend's own lifecycle applies | Container | Mar 2026 | Python 3.10+ SDK (control plane 3.11+), PyPI bespokelabs-sandbox 0.4.0 (2026-11-11)Python 3.10+ SDK (control plane 3.11+), PyPI bespokelabs-sandbox 0.4.0 (2026-11-11); one API with explicit backend selection; README says automatic failover and cheapest-provider routing are not implemented, though the PyPI summary calls it "OpenRouter for Sandboxes"; self-hosted control plane with dashboard: per-sandbox CPU/RAM/GPU charts, compute spend and model token costs, organization-scoped API keys, concurrency, spend, GPU and lifetime policies, audit log, CSV export; preset images with Claude Code and Codex CLI; run_agent for hosted agent runs; network policy; port exposure; detached jobs on Google Cloud Batch; by Bespoke Labs; 6 GitHub stars |
| Unified API | Unified API | Provider's cost with no markup; BYOK: you pay providers directly, no BoxRouter feeProvider's cost with no markup; BYOK: you pay providers directly, no BoxRouter fee; BoxRouter Credits carry a 5% processing fee on purchase | Depends on provider (storage on providers that charge for a paused sandbox) | Destroyed at the end of lifetime_seconds by default, or paused with at_end: pauseDestroyed at the end of lifetime_seconds by default, or paused with at_end: pause; optional idle_shutdown requirement uses the route's own idle shutdown; what a pause keeps depends on the provider | — | Oct 2026 | One API and SDKs (Python, TypeScript) over several sandbox providersOne API and SDKs (Python, TypeScript) over several sandbox providers; routes by price or run time, or to providers you pick, and fails over when a provider refuses a sandbox; benchmarks page; GitHub Actions jobs and agent deployments on routed sandboxes; volumes; spend caps |
| Unified API | Unified API | Pay as you go: 5% platform fee on credit purchasesPay as you go: 5% platform fee on credit purchases, then each provider's listed compute rate (the rate table showed 'Rates temporarily unavailable' on 2026-11-14); $5 free credit; bring your own provider keys and the provider bills you; Enterprise custom | — | — | — | Jul 2025 | Sandbox Router over providers including E2B, Isorun, Superserve, Novita, Modal, Tensorlake, Mosaic and FreestyleSandbox Router over providers including E2B, Isorun, Superserve, Novita, Modal, Tensorlake, Mosaic and Freestyle; Actions runs GitHub workflows; public provider benchmarks; open-source SDK with about 70 provider packages (MIT, 292 GitHub stars); 'Managed agents' coming soon; run by Snelling, LLC |
| Unified API | Unified API | AutoPilot was 'Free during preview' in the 2026-11-05 release that added itAutoPilot was 'Free during preview' in the 2026-11-05 release that added it; sandboxes run in your own provider accounts and each provider bills you directly | Depends on provider | Depends on the provider; CostGraph only records when each sandbox starts and stops | — | Nov 2026 | Router and cost tracker over sandbox providers: wraps ComputeSDK providers with @costgraph/computesdk (TypeScript) or aRouter and cost tracker over sandbox providers: wraps ComputeSDK providers with @costgraph/computesdk (TypeScript) or a REST API for other languages; routing policy cheapest, balanced or fastest (start times from ComputeSDK's public benchmark), off by default; cost per agent, run and customer from metadata tags; can also start sandboxes on provider accounts connected to CostGraph; part of the CostGraph cloud-cost platform (Kubernetes, CI, GPUs) |
| Unified API | Unified API | Free, open source (MIT); Crabbox does not meter or resell compute and has no hosted planFree, open source (MIT); Crabbox does not meter or resell compute and has no hosted plan or markup; you pay the selected provider and your own coordinator infrastructure (Cloudflare, or Node.js with PostgreSQL); crabbox marketplace quotes are preview-only (price per hour for 2 vCPU / 4 GiB: n/a (free software; you pay the chosen provider)) | N/a (the provider's own rates) | A lease holds a box until it is released or expiresA lease holds a box until it is released or expires; coordinator defaults are a 30-minute idle timeout and a 30-minute TTL, with a 4-hour (14,400 s) policy maximum; crabbox run auto-extends by 30 minutes while a command runs; a one-shot run releases the box after the command, --keep or warmup keeps a warm box; pause, resume and checkpoints where the provider supports them | — | Apr 2026 | CLI that leases a remote box, syncs your uncommitted checkout, runs the command and streams output backCLI that leases a remote box, syncs your uncommitted checkout, runs the command and streams output back; 85 provider catalog entries (84 built in): clouds, delegated sandboxes such as E2B, Daytona, Modal, Vercel Sandbox and Upstash Box, local VMs, self-hosted microVM, GPU clouds; Linux, macOS, native Windows and WSL2 targets; VNC/WebVNC desktops, a browser and VS Code in the browser on managed Linux; optional self-hosted coordinator with lease TTLs, spend caps and usage tracking; checkpoints; GitHub openclaw/crabbox, 1,538 stars; install with brew install openclaw/tap/crabbox |
| Unified API | Unified API | Free, open source (Apache-2.0), self-hosted; no hosted servicePricing page | — | From OpenSandbox: pause/resume on Docker and Fast Sandbox (memory kept)From OpenSandbox: pause/resume on Docker and Fast Sandbox (memory kept), stop/start on Kubernetes, provider timeout as a TTL backstop; a paused microVM returns its pool slot | Container | Oct 2026 | Self-hosted control plane over sandbox backends: multi-tenancy, quotas and fair queueing, policy-gated placementSelf-hosted control plane over sandbox backends: multi-tenancy, quotas and fair queueing, policy-gated placement, tamper-evident audit log; adapters for OpenSandbox, Daytona and Freestyle plus 13 experimental (E2B, Vercel, AWS, Azure, Google Cloud...); SDKs for TypeScript, Python and Go plus a CLI; Helm chart, air-gapped install; 7 GitHub stars |
| Unified API | Unified API | Free, open source (Apache-2.0); you use your own provider accountsPricing page | Billed by the chosen provider | Depends on the provider: snapshot on Sprites, Modal, Docker, Daytona and BlaxelDepends on the provider: snapshot on Sprites, Modal, Docker, Daytona and Blaxel; unarchive on Blaxel; restart from disk on Sprites, E2B, Docker and Daytona | Container | Oct 2026 | One Go CLI over six sandbox providers: create, list, exec, console, sync (copy a repository in), restart, unarchiveOne Go CLI over six sandbox providers: create, list, exec, console, sync (copy a repository in), restart, unarchive, upgrade, snapshot, url, destroy, and 'mcp' to serve sandboxes to agents; Homebrew, install script or go install; the Docker provider is experimental and needs a Docker Agentic Platform subscription; 5 GitHub stars |
| Unified API | Unified API | Free, open source (Deep Agents, MIT); each provider bills its own sandboxesFree, open source (Deep Agents, MIT); each provider bills its own sandboxes. LangChain's own hosted LangSmith Sandboxes are the LangSmith Sandboxes row | Billed by the chosen provider | Depends on the provider; an app chooses one sandbox per thread or one shared per assistant | — | Nov 2025 | In Deep Agents, sandboxes are backends that give the agent an execute tool plus file tools (ls, read_file, write_fileIn Deep Agents, sandboxes are backends that give the agent an execute tool plus file tools (ls, read_file, write_file, edit_file, glob, grep); 27 listed integrations including Daytona, Modal, E2B, Runloop, Vercel, AWS AgentCore, Databricks, OpenSandbox, NVIDIA OpenShell, Kubernetes, Islo, Upstash Box, boxd, NoMac and LangSmith; Deep Agents Code has a --sandbox flag; deepagents repo 31,287 GitHub stars |
| Unified API | Unified API | SDK free (MIT); sandbox compute is billed by the chosen providerSDK free (MIT); sandbox compute is billed by the chosen provider. Code Interpreter and Hosted Shell containers: 1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session, eligible sessions billed per minute with a 5-minute minimum (OpenAI (Codex cloud, Agents API sandboxes) and OpenAI Containers (Code Interpreter / Hosted Shell) cover these containers) | Set by the chosen provider | Sandbox Agents: lifecycle set by the chosen providerSandbox Agents: lifecycle set by the chosen provider; RunState, serialized session state and workspace snapshots let later runs resume a session or seed a fresh one. Code Interpreter: a container expires after 20 minutes unused and its data is discarded | Container | Apr 2026 | SandboxAgent with a Manifest (files, directories, Git repos, environment, users, groups, mounts), capabilities (shellSandboxAgent with a Manifest (files, directories, Git repos, environment, users, groups, mounts), capabilities (shell, filesystem editing, image inspection, skills, memory, compaction), exposed ports and mounts for S3, Cloudflare R2, GCS and Azure Blob; Python and TypeScript Agents SDKs (openai-agents-python: MIT, 30,685 GitHub stars); harness stays in your app while compute runs in the sandbox |
| Unified API | Unified API | Self-host free (Apache-2.0). Hosted, BYOK for models and charging only sandbox computeSelf-host free (Apache-2.0). Hosted, BYOK for models and charging only sandbox compute: Free $0 with a $1 one-time trial credit; Pro $20/mo for $22 credit ('~73 hours cloud sandbox'); Max $100/mo for $120 credit ('~400 hours'), about $0.30 per sandbox-hour, but the sandbox size is not stated | Hosted: not stated; self-hosted: billed by the chosen provider | Depends on the driver: E2B suspend/resume with a memory snapshot, Daytona stop/startDepends on the driver: E2B suspend/resume with a memory snapshot, Daytona stop/start, Sprites automatic pause/wake, Vercel snapshot/restore; node-docker and Cloudflare cannot suspend; OpenMA keeps session history and checkpoints across sandboxes | Container | Apr 2026 | Open-source alternative to Claude Managed Agents and the OpenAI Agents API, compatible with both SDKsOpen-source alternative to Claude Managed Agents and the OpenAI Agents API, compatible with both SDKs; harnesses: default, Pi and ACP adapters (Codex, Claude Code, OpenCode); vaults with a credential proxy; skills, memory, MCP; Backchat desktop app, OMA CLI and TypeScript SDK; runs on Node + Docker, Cloudflare, Fly.io or Render; 360 GitHub stars |
| Unified API | Unified API | Managed capacity is drawn from prepaid organization credits at each provider's usageManaged capacity is drawn from prepaid organization credits at each provider's usage; 5.5% fee on credit purchases (minimum $0.80); one-time $500 welcome credit; BYOK capacity is billed by the provider and uses no credits | Billed at the chosen provider's rates | Create, pause, resume and destroy through one lifecycle APICreate, pause, resume and destroy through one lifecycle API, with pause only on providers that support it; managed sandboxes are scheduled to stop when the balance reaches zero | VM | Aug 2026 | Compute gateway: REST API, TypeScript SDK, CLI and Agent Skill for provisioning CPU sandboxes across providersCompute gateway: REST API, TypeScript SDK, CLI and Agent Skill for provisioning CPU sandboxes across providers, with capability-based routing, safe fallback, pinned providers and one usage report by day, provider, project and sandbox; argv processes, file transfer and time-limited HTTP endpoints, but no interactive terminals, SSH or WebSockets; webhooks; GPU workloads, browser computers and persistent machines are planned, not available; open source (MIT), 4 GitHub stars |
| Unified API | Unified API | Free, open source; self-hosted on your own serversFree, open source; self-hosted on your own servers, so infrastructure costs are yours (Apache-2.0); hosted providers such as E2B bill separately | — | Sessions keep one sandbox and its files for many callsSessions keep one sandbox and its files for many calls; warm pools keep sandboxes ready to claim; idle policy not stated | WebAssembly | Sep 2026 | Sandbox service (plimsolld) that runs agent-written code across providersSandbox service (plimsolld) that runs agent-written code across providers; callers can require a minimum isolation tier per request; signed run records with content hashes; credential broker with per-run API grants so guest code never holds keys; tools generated from an OpenAPI document; Go library plus Python client; Claude Managed Agents self-hosted worker adapter; 28 stars |
| Unified API | Unified API | No prices published; the managed platform needs no API key for the default CloudflareNo prices published; the managed platform needs no API key for the default Cloudflare provider, and other providers use your own API keys | — | Sandboxes are ephemeral by default; stop, start and delete from the CLISandboxes are ephemeral by default; stop, start and delete from the CLI; volumes, files and secrets persist across sessions; point-in-time backups of directories | — | Feb 2024 | One CLI, API and SDK over 10 providers: Cloudflare, Daytona, Vercel, E2B, Modal, Runloop, HopX, Deno DeployOne CLI, API and SDK over 10 providers: Cloudflare, Daytona, Vercel, E2B, Modal, Runloop, HopX, Deno Deploy, Fly Sprites, Blaxel; sandbox definitions stored as AT Protocol records, sign-in with a Bluesky handle; SDKs for JavaScript/TypeScript, Gleam, Clojure and Elixir (Rust and Go repos too); SSH keys, secrets, preview URLs, VS Code in the browser, services; works with Claude Code, Codex CLI, Gemini CLI, Copilot, OpenClaw; open source (MPL-2.0), 73 GitHub stars; Bluesky @pocketenv.io, no X link |
| Unified API | Unified API | The provider's raw per-second cost, no markupThe provider's raw per-second cost, no markup; one fee at top-up: 5.0% for USDC over x402 on Base, 5.5% (minimum $0.80) by card; $1 minimum top-up; card accounts start with $5 free credits; the docs' example 2 vCPU / 4 GB / 8 GB machine routed to Daytona at $0.165924/h | N/a (no stopped state; machines are destroyed) | No pause: a machine is destroyed when ttl_seconds expires (default 4 hNo pause: a machine is destroyed when ttl_seconds expires (default 4 h; some offers cap it lower), when accrued spend reaches max_spend, on DELETE, or when the account balance reaches zero | VM | Sep 2026 | 'OpenRouter for sandboxes': ask for CPU (1-8), RAM (1-16 GB), disk and region (us-west, us-east'OpenRouter for sandboxes': ask for CPU (1-8), RAM (1-16 GB), disk and region (us-west, us-east, eu-central) and get the cheapest fitting offer; ssh access, HTTP exec (default 300 s, max 600 s) or both; expose HTTP/WebSocket ports through the provider's own bridge URL; max_rate and max_spend filters; idempotent client_ref; no-account use where the paying wallet becomes the account; agent skill file (skill.md v1.4.0); part of reachpad, an app publishing service |
| Unified API | Unified API | SDK free, open source (MIT in package.jsonSDK free, open source (MIT in package.json; the repo has no LICENSE file). Sandbank Cloud: $0.02 per sandbox for 10 minutes and $0.02 per 10-minute extension (= $0.12/hour), exec, files and proxy free; paid in USDC on Base via x402 with no signup, or with an API token; sandbox limit 2 vCPU / 1,024 MB RAM / 5 GB disk | N/a (Cloud sandboxes are destroyed, not paused) | Sandbank Cloud: 10-minute sessions, extended 10 minutes at a time, destroyed when doneSandbank Cloud: 10-minute sessions, extended 10 minutes at a time, destroyed when done; snapshot support in the Cloud, Cloudflare and BoxLite adapters; the Chrome browserbox can save and restore VM snapshots | VM | Mar 2026 | One TypeScript interface (@sandbank.dev/core) with adapters for Sandbank Cloud, Daytona, Fly.ioOne TypeScript interface (@sandbank.dev/core) with adapters for Sandbank Cloud, Daytona, Fly.io, Cloudflare and BoxLite, plus capability detection (streaming, port expose, snapshot, sleep); multi-agent sessions with relay and shared context; DB9 serverless Postgres service adapter; Sandbank Cloud's default codebox image has Claude Code and Codex, a ttyd web terminal, warm pool for sub-second starts, a Lightpanda browser in every sandbox and an optional Chrome browserbox micro-VM; lifecycle webhooks; site shows v0.2.0 while npm's latest is 0.3.6; GitHub chekusu/sandbank 179 stars; by chekusu, Inc. |
| Unified API | Unified API | Sandboxes $0, 'free for now'; fair-use caps may applySandboxes $0, 'free for now'; fair-use caps may apply; pricing to be announced before it changes; no CPU/RAM size published. The rest of SandBase (models, data APIs) is pay as you go | — | Created and stopped through the API with a task timeoutCreated and stopped through the API with a task timeout; pause and resume 'where supported' by the provider; Playground notes that paused instances may retain resources and running instances follow the server timeout | — | — | E2B-compatible gateway: official E2B JS/Python SDK 2.51.0 with a SandBase key and E2B_API_URL=sandbox.sandbase.aiE2B-compatible gateway: official E2B JS/Python SDK 2.51.0 with a SandBase key and E2B_API_URL=sandbox.sandbase.ai; commands, files, public or org-private templates; Sandbox Console with status, logs and metrics; experimental Playground (commands up to 25 s, no snapshots or fork); separate from SandBase's hosted Agent Sessions; one key also covers SandBase's model and API marketplace |
| Unified API | Unified API | Free, open source (Apache-2.0)Pricing page | — | — | Container | Jan 2026 | Static-binary server that runs inside a sandbox and exposes HTTP/SSE, so an app can control Claude Code, CodexStatic-binary server that runs inside a sandbox and exposes HTTP/SSE, so an app can control Claude Code, Codex, OpenCode, Amp or Pi through one API; universal session and event schema, persisted to Postgres or Rivet; TypeScript SDK with sandbox provider adapters; OpenCode support experimental; repo rivet-dev/sandbox-agent now redirects to rivet-dev/agents (1,645 GitHub stars), which also holds Rivet Agents |
| Unified API | Unified API | $0.2448 | Depends on the provider; not stated for stopped Standby Pool sandboxes | Auto-stop after idleStopMinutes without activity (5-240)Auto-stop after idleStopMinutes without activity (5-240); start, stop and restart keep the sandbox, and what survives a stop depends on the provider; provider snapshots on request; exec timeout 600 s by default, up to 7,200 s (86,400 s detached); delete removes the sandbox and its provider snapshots | VM | Nov 2026 | One API, CLI and console across sandbox providers; Auto routing fails over to the next eligible providerOne API, CLI and console across sandbox providers; Auto routing fails over to the next eligible provider, then the Standby Pool; explicit provider order and pool labels for provider, region, isolation or security requirements; public provider catalog comparing price, isolation, snapshots, persistence and limits with ComputeSDK benchmark results; runs triggered from GitHub, Linear, webhooks or the API; hosted MCP servers behind an MCP gateway; connectors (GitHub) and egress allowlists; container registry at registry.caas.sh; schedules; the CLI package @sandboxio/cli is not on npm yet; no open-source repo found |
| Unified API | Unified API | Free, open source (MIT); each provider bills its own sandboxesPricing page | Billed by the chosen provider | Depends on the adapter; adapters declare snapshot create, delete and restore supportDepends on the adapter; adapters declare snapshot create, delete and restore support; withSandbox stops the sandbox after the callback | — | May 2026 | Unified TypeScript API (@sandbox-sdk/core) for files, processes, ports and snapshots, with one package per adapterUnified TypeScript API (@sandbox-sdk/core) for files, processes, ports and snapshots, with one package per adapter, capability checks, a typed sandbox.raw escape hatch to the native client and normalized SandboxError codes; AI SDK and Claude tool definitions; runs on Node, Bun and Workers; 23 GitHub stars |
| Unified API | Unified API | Free, open source (MIT); each provider bills its own sandboxesPricing page | Billed by the chosen provider | Depends on the adapter: create, reconnect by id, pauseDepends on the adapter: create, reconnect by id, pause, snapshot or destroy where supported; 'await using' deletes the sandbox when the scope exits | Container | Jul 2026 | One TypeScript package for commands, files and git across 13 providersOne TypeScript package for commands, files and git across 13 providers; every adapter passes the same conformance suite; non-zero exits returned as ok: false; git tokens kept out of logs; in-memory adapter for CI with no keys; 115 GitHub stars |
| Unified API | Unified API | Free, open source (MIT) library and CLI; you pay each underlying provider at its own ratesPricing page | N/a (library; each provider's own idle pricing applies) | Run() and run_many() create a temporary sandbox and destroy it afterwardsRun() and run_many() create a temporary sandbox and destroy it afterwards; the async context manager cleans up on exit; the CLI's --keep keeps a sandbox and labels with get_or_create/--reuse reattach to it; per-command timeout (-t); named persistent sandboxes for `sandboxes claude` on Sprites only; otherwise the provider's own lifecycle applies | Container | Sep 2025 | Python 3.11+ library (PyPI cased-sandboxes 0.7.0, 2026-02-17) plus CLIPython 3.11+ library (PyPI cased-sandboxes 0.7.0, 2026-02-17) plus CLI; one API over E2B, Modal, Daytona, Hopx, Vercel and Sprites (Fly.io), Cloudflare experimental via a self-hosted Worker; auto-selects the first provider with credentials; automatic failover through a fallback list; label-based sandbox reuse; upload/download and streaming output; CLI runs Python, Node, TypeScript, Go, Rust and bash files with language auto-detect; `sandboxes claude` starts Claude Code in a cloud sandbox (Sprites recommended); 108 GitHub stars |
| Unified API | Unified API | Free, open source (MIT) library; you pay each provider at its own ratesPricing page | N/a (library; each provider's own idle pricing applies) | Suspend() and resume() methods; the README's matrix marks memory persistence on E2BSuspend() and resume() methods; the README's matrix marks memory persistence on E2B, Beam and CodeSandbox and file persistence on all five; destroy(); otherwise the provider's own lifecycle applies | — | Jun 2025 | TypeScript/JavaScript library on npm (@gitwit/sandbox 1.0.3, last published 2025-09-24)TypeScript/JavaScript library on npm (@gitwit/sandbox 1.0.3, last published 2025-09-24); one interface over E2B, Daytona, Beam, CodeSandbox and Modal; runCommand with background processes, file read/write/list/move/delete, preview URLs, pseudo-terminals, env vars; buildTemplate turns a Dockerfile into an E2B template or a Daytona snapshot; per-provider feature matrix; by James Murdza; 34 GitHub stars; the input's repo synacktraa/sandboxjs is a 0-star fork of this one and lacks Beam |
| Unified API | Unified API | Open source with your own provider keys: free, you pay each providerOpen source with your own provider keys: free, you pay each provider; one hosted sandboxRouter key: provider list price + 5.5% | — | — | — | Nov 2026 | One API that starts each sandbox on the cheapest provider that fits and falls back when one failsOne API that starts each sandbox on the cheapest provider that fits and falls back when one fails; 33 providers listed, 18 pass its real-account check; TypeScript SDK @sandboxrouter/sdk, CLI and HTTP API; agents can sign up for a key themselves; MIT, v0.8.3, 0 GitHub stars; hosted keys run by AlgiAI, Corp. |
| Unified API | Unified API | Free, open source (MIT)Pricing page | N/a (self-hosted or billed by the chosen provider) | Sandboxes created per use with sandbox and execution timeouts; no suspend model | Container | Jan 2026 | Go library for running AI-generated code through one API on any of 11 providers, with automatic language detectionGo library for running AI-generated code through one API on any of 11 providers, with automatic language detection, per-language runtimes and timeouts; maintained by Happy Hacking Space, a nonprofit hacker community, whose site (the repo's homepage) does not mention sindoq; 9 GitHub stars |
| Unified API | Unified API | Not published; an org connects its own provider accounts in the consoleNot published; an org connects its own provider accounts in the console, and stackbox shows estimated cost per sandbox and provider from the providers' list prices; --max-price caps the price per sandbox-hour | — | Lifetime set per sandbox with --timeout (default 30 minutesLifetime set per sandbox with --timeout (default 30 minutes; providers whose limit is shorter are skipped); a sandbox keeps running between commands until it is deleted or times out; --rm deletes it after the command | Container | Nov 2026 | Hosted router that puts E2B, Modal, Daytona, Blaxel and Cloudflare Containers behind one OpenSandbox-compatible API andHosted router that puts E2B, Modal, Daytona, Blaxel and Cloudflare Containers behind one OpenSandbox-compatible API and one key (OpenSandbox SDKs, osb and Harbor's opensandbox environment work unchanged against api.box.stacktrace.ai); placement strategies priority, cost, latency, balanced, random, round-robin; per-provider image preparation and warm pools; dry-run shows where a sandbox would go; usage and estimated cost per provider; Cloudflare bridge deployed on your account; Python CLI (Apache-2.0; the hosted service is proprietary); org API keys |
| Unified API | Unified API | Free, open source (MIT); you pay the sandbox provider and model provider you choosePricing page | N/a (provider's pricing) | SDK exposes pause(), resume() and kill() on the provider sandbox (pause documented forSDK exposes pause(), resume() and kill() on the provider sandbox (pause documented for the Codex agent); idle behavior otherwise follows the provider | Container | May 2025 | Runs Claude Code, Codex, Gemini CLI, Grok CLI, OpenCode and other coding agents in a sandboxRuns Claude Code, Codex, Gemini CLI, Grok CLI, OpenCode and other coding agents in a sandbox, with secret redaction and observability (logs, traces, metrics); CLI 'vibekit claude' works offline; TypeScript SDK @vibe-kit/sdk (last npm release 0.0.70 on 2025-10-14) with GitHub PR, branch and test helpers; @vibe-kit/auth for using MAX subscriptions in apps; by Superagent Technologies; vibekit.sh returned Vercel DEPLOYMENT_NOT_FOUND on 2026-11-14, docs.vibekit.sh still up; GitHub 1,867 stars, last push 2026-01-13 |
| Unified API | Unified API | Free, open source (Apache-2.0); Vercel or Cloudflare bill the sandboxesPricing page | Billed by Vercel or Cloudflare | Each sandbox task run starts a new sandbox, with a timeout set in the task's package.json | Container | Apr 2026 | Sandbox is one of about 30 server primitives in a Vite plugin (KV, database, blob, queue, workflowSandbox is one of about 30 server primitives in a Vite plugin (KV, database, blob, queue, workflow, agents and more) that keep one API across hosts; sandbox tasks are folders under server/sandbox-tasks; Agents with Codex, Claude Code or AI SDK drivers and GitHub channels; 31 GitHub stars |
| Dev environment | Dev environment | $0.0685 | Free: $0 while asleep; paid plans are flat monthly (same as running) | Free plan's preview sleeps after 4 hours without activity and wakes when you open theFree plan's preview sleeps after 4 hours without activity and wakes when you open the workspace or preview URL (code and data kept); paid plans are always on | Container | Sep 2026 | Site now leads with 'cloud hosting for apps built with AI'Site now leads with 'cloud hosting for apps built with AI'; its /sandboxes page sells an agent workspace where Claude Code, Codex, Cursor, ChatGPT or other MCP clients edit files, run commands, read logs and open previews through the hosted MCP server (OAuth); production deploys need your approval; managed Postgres, pgvector, Redis, MongoDB, MySQL, RabbitMQ; S3-compatible buckets and persistent volumes; custom domains and TLS; adios.yaml manifest; CLI and Git deploys; workflows (cron, webhooks, approvals); Next.js, Node.js, Python, Go, PHP, Ruby, .NET |
| Dev environment | Dev environment | Self-serve tiers, prices billed annually (monthly billing in brackets): Basic $18 / $36 /Self-serve tiers, prices billed annually (monthly billing in brackets): Basic $18 / $36 / $54 per month ($20 / $40 / $60) with 480 / 960 / 1,440 included minutes per OS and 1 / 3 / 5 concurrent VMs, macOS M2 Pro Large 6 vCPU / 14 GB, Linux up to Medium 4 vCPU / 16 GB; Pro $36 / $72 / $108 ($40 / $80 / $120), macOS M2 Pro X Large 12 vCPU / 28 GB, Linux up to Large 8 vCPU / 32 GB; overage shown on the Basic card: macOS $0.0288/min, Linux $0.0044/min ($0.264/h), up to a monthly cap (Basic macOS $99 / $297 / $495, Linux $39 / $117 / $195; Pro macOS $199 / $597 / $995, Linux $79 / $237 / $395), after which usage is free for the month; 14-day trial; Enterprise priced on committed capacity. No 2 vCPU / 4 GiB machine is listed | $0 compute when terminated; session storage 'isn't billed separately today'$0 compute when terminated; session storage 'isn't billed separately today' (FAQ) | Auto-terminate timer after 5 days of inactivity by default (settable up to 30 days orAuto-terminate timer after 5 days of inactivity by default (settable up to 30 days or off; free trial capped at 8 hours); terminating stops the machine and keeps the persistent disk (a diff against the stack image) until you delete it or the stack is removed; restore takes about 20 s to a few minutes; warm pools (alpha) keep booted machines ready | VM | Jun 2026 | FAQ: Linux from 2 to 48 vCPUs, macOS on M2 Pro, M4, M4 ProFAQ: Linux from 2 to 48 vCPUs, macOS on M2 Pro, M4, M4 Pro; iOS simulator and Android emulator device sessions that agents screenshot, click and type on; shareable device preview links; MCP server (MIT) at mcp-rde.bitrise.io; 'bitrise-cli rde claude' opens Claude Code on a fresh session; HTTP API (experimental); VS Code extension; SSH; VNC desktop on macOS only; Docker on Linux, no nested virtualization on macOS; templates with lifecycle hooks; short-lived GitHub tokens injected; OIDC to AWS and GCP; same stacks as Bitrise CI; dedicated or private cloud on Enterprise; coming soon: default-deny egress, host-side secrets, custom VM images |
| Dev environment | Dev environment | Per-user plans with included box-hours (1 box-hour = 1 h awake on the standard 4 vCPU / 8Per-user plans with included box-hours (1 box-hour = 1 h awake on the standard 4 vCPU / 8 GiB / 100 GB box): Trial $0 with 20 box-hours; Starter $19/user/mo with 50; Pro $99/user/mo with 300; Max $249/user/mo with 750; extra usage $0.40/box-hour (prepaid credit, 5-10% off); larger sizes use more box-hours (8 vCPU/16 GiB = 2 box-hours/hr, $0.80); no 2 vCPU / 4 GiB size; billed by the second; Teams custom | $0 (sleeping devboxes are free) | Idle devboxes sleep automatically; sleeping devboxes are free and unlimitedIdle devboxes sleep automatically; sleeping devboxes are free and unlimited; devboxes start in seconds from a Template box snapshot of filesystem and memory | — | Jun 2026 | A cloud devbox per coding agent; runs native Claude Code and Codex with your own subscription or API key (no resoldA cloud devbox per coding agent; runs native Claude Code and Codex with your own subscription or API key (no resold tokens); start agents from Slack, Linear, schedules or API; desktop, iOS, Android, web and CLI clients; SSH and IDE-over-SSH; public preview URLs per port; agent migrates your local dev setup to a Template box; Teams: SSO/SCIM, audit logs, BYO cloud/VPC; taking over Conductor Cloud users (shut down 2026-11-20) |
| Dev environment | Dev environment | EUR only: Hobby €7/month (1 active environment, 2 CPU cores, 4 GB RAMEUR only: Hobby €7/month (1 active environment, 2 CPU cores, 4 GB RAM; 2-day trial) = €0.0096/h (€7/month ÷ 730); Growth €20/month (5 active environments at 2 CPU / 4 GB each, extra environments €5/month); Team €39/month (5 environments, 5 members); Enterprise custom CPU and RAM; yearly billing saves up to 20%; bring your own AI subscription or key | Same as running (flat monthly plan; environments stay up) | Environments keep running when you close your laptop: terminal sessionsEnvironments keep running when you close your laptop: terminal sessions, running processes and file changes persist; no idle suspend or timeout stated | Container | Jun 2025 | Hosted cloud dev environments for AI coding agents with Claude Code, Cursor CLI, CodexHosted cloud dev environments for AI coding agents with Claude Code, Cursor CLI, Codex, Gemini CLI and OpenCode preinstalled; web and mobile UI (PWA) with chat, terminal, file explorer and Git; continue the same session in VS Code or Cursor over SSH; REST API, webhooks and an n8n node; start sessions from Linear or Jira; shared team environments, MCP servers and context files; encrypted keys; self-hostable open-source UI (siteboon/claudecodeui, formerly Claude Code UI, AGPL-3.0, 14,617 GitHub stars) |
| Dev environment | Dev environment | Community: free, open source (AGPL-3.0), self-hosted on your own infrastructureCommunity: free, open source (AGPL-3.0), self-hosted on your own infrastructure; Premium: priced annually per user, contact sales; compute is billed by your own cloud | A stopped workspace costs whatever your own cloud charges for what the templateA stopped workspace costs whatever your own cloud charges for what the template keeps, such as disks | Workspaces stop after an inactivity timeout set per template (activity from SSH, IDEWorkspaces stop after an inactivity timeout set per template (activity from SSH, IDE, web terminal and other sessions bumps it); autostart and autostop schedules; Premium adds a required periodic stop, user quiet hours, and dormancy, which marks long-unused workspaces dormant and then deletes them | Container | Dec 2021 | Self-hosted cloud development environments for developers and coding agents (Claude Code, CodexSelf-hosted cloud development environments for developers and coding agents (Claude Code, Codex, Cursor and others in ephemeral workspaces); Coder Agents (beta 2026-05-06): self-hosted agent loop in the Coder control plane with chat UI and Agents API, LLM keys kept out of workspaces, workspaces provisioned only when compute is needed; AI Governance gateway; WireGuard tunnel; Linux, Windows and macOS, ARM and x86-64; VS Code, JetBrains, Cursor and Jupyter; VPC, on-prem or air-gapped; GitHub coder/coder 17,384 stars |
| Dev environment | Dev environment | $0.2096 | $0 compute when suspended; volume storage $0.21224989/GB-month (General Purpose$0 compute when suspended; volume storage $0.21224989/GB-month (General Purpose SSD) | Suspend scales the workload to zero and keeps the /root volumeset (storage billed only)Suspend scales the workload to zero and keeps the /root volumeset (storage billed only); resume remounts it; delete removes both; no automatic idle suspend described | Container | Jul 2026 | Ready in 60-90 s; browser IDE (code-server) and terminal (ttyd), or VS Code, CursorReady in 60-90 s; browser IDE (code-server) and terminal (ttyd), or VS Code, Cursor, JetBrains or SSH through 'cpln sandbox connect'; templates or your own OCI image; an identity and secrets attached at creation let the agent reach private services through Universal Cloud Identity and Cloud Wormhole with no keys in its context; runs on Control Plane's AWS, GCP, Azure and OCI accounts, your cloud or bare metal; tmux sessions; MCP server for the platform; GPUs on the platform |
| Dev environment | Dev environment | Contact sales for every plan: Starter (per seat), Business (annual contract)Contact sales for every plan: Starter (per seat), Business (annual contract), Enterprise (custom terms, on-prem, multi-cloud); no list prices | — | Auto-suspends idle sandboxes after an org-level periodAuto-suspends idle sandboxes after an org-level period; idle detection counts SSH, exec, port-forward, remote desktop, IDE and endpoint connections and running AI agents, plus custom probes and statistical 'enhanced' detection; pinned sandboxes skip suspension up to an admin-set maximum; Mission Critical state blocks suspension and rescheduling; on_suspend, on_resume and on_delete hooks | Container | Mar 2026 | Production-like agent sandboxes in your own cloud: database snapshots, internal servicesProduction-like agent sandboxes in your own cloud: database snapshots, internal services, caches and queues scoped per agent; network isolation, with credentials injected only when needed and never shown to the model; SOC 2 Type II; native support for Claude Code, Cursor agents and Codex; built-in agent harness with multi-agent teams and your own model key; Kubernetes traffic interception against a shared staging tier; Web IDE, SSH, JetBrains and remote desktop; Mac sandboxes on EC2 Mac; workspace snapshots; Crafting Express (GA 2026-09-28) runs the platform on a laptop in Docker; multi-region and multi-cloud |
| Dev environment | Dev environment | $0.18 | Same as running (compute is metered until the environment is deleted) | One machine per branch, metered from creation until deletionOne machine per branch, metered from creation until deletion; reaped when the pull request merges; no stop or suspend described; the agent runs in tmux on the instance and keeps running when the browser closes | VM | Sep 2026 | Disposable cloud dev environment per branch with Claude Code preinstalled (Codex and Gemini CLI 'coming soon')Disposable cloud dev environment per branch with Claude Code preinstalled (Codex and Gemini CLI 'coming soon'), using your own Claude subscription or API key with no proxying; root on the machine; browser terminal, SSH via ssh.ephemerals.dev and a preview URL per port on a separate domain; secrets sealed with KMS and deleted with the branch; environments owned by the organization; mobile workspace and Slack app in progress |
| Dev environment | Dev environment | Compute per hour: 2-core $0.18, 4-core $0.36, 8-core $0.72, 16-core $1.44, 32-core $2.88Compute per hour: 2-core $0.18, 4-core $0.36, 8-core $0.72, 16-core $1.44, 32-core $2.88; storage $0.07/GB-month; machine types 2 cores/8 GB RAM, 4/16, 8/32, 16/64, 32/128 (pricing calculator); personal accounts include 120 core-hours and 15 GB-month (Free) or 180 core-hours and 20 GB-month (Pro); no free quota for organizations. No 2 vCPU / 4 GiB size | $0 compute; storage $0.07/GB-month | Stops after an idle timeout (default 30 min, settable 5 to 240 min)Stops after an idle timeout (default 30 min, settable 5 to 240 min); 12 h maximum lifetime; a stopped codespace keeps its files and is deleted after 30 days of inactivity by default (retention configurable) | VM | Aug 2021 | Cloud dev environments from dev container configs; browser, VS Code or JetBrains; prebuildsCloud dev environments from dev container configs; browser, VS Code or JetBrains; prebuilds; created from the web or GitHub CLI; isolated virtual network per codespace with inbound blocked; the pages read do not market Codespaces as an agent sandbox |
| Dev environment | Dev environment | No price: sandboxes burn core-hours that come from 'platform gifts plus activityNo price: sandboxes burn core-hours that come from 'platform gifts plus activity earnings', not direct billing (resource-pool doc); capacity errors read 'DevKit free resources are busy'; 1 sandbox per pool type per user, 5 concurrent per account; the plugin itself is Apache-2.0 | No price; a shut-down persistent sandbox keeps its data for up to 7 days | Two pools: ephemeral (default) shuts down after 30 min idle and is deleted after 1 hTwo pools: ephemeral (default) shuts down after 30 min idle and is deleted after 1 h idle; persistent (32 GiB EVS data disk) shuts down after 1 h idle, keeps the instance and data, and is reclaimed after at most 7 days idle; liveness is detected only on connect; the server echoes these values and they may change | Container | Jul 2026 | Official Huawei Cloud plugin for coding agents (OpenCode, Codex CLI and desktop, CodeArts Agent and Work, WorkBuddyOfficial Huawei Cloud plugin for coding agents (OpenCode, Codex CLI and desktop, CodeArts Agent and Work, WorkBuddy, DeepSeek Harness, OfficeAce, Hermes, OpenClaw, AtomCode) with skills, MCP tools and guardrails for 20+ Huawei Cloud services; its DevStation sandbox is a temporary runtime for deploying and previewing web apps, with git upload, remote exec, nginx helpers and public tunnel URLs (1-720 h, default 72 h, renewable); pool flavor 3 vCPU / 6 GB (resource-pool doc, 2026-11); telemetry on by default; 71 GitHub stars |
| Dev environment | Dev environment | Devbox Minute $0.004; Linux S (up to 4 vCPU, 8 GB) 1 Devbox Minute/min = $0.24/hrDevbox Minute $0.004; Linux S (up to 4 vCPU, 8 GB) 1 Devbox Minute/min = $0.24/hr, M (8 vCPU, 16 GB) $0.48/hr, L (16/32 GB) $0.96/hr, XL (32/64 GB) $1.92/hr; macOS M (6 vCPU, 14 GB) $0.06/min, L (12/28 GB) $0.12/min; storage $0.20/GB-month of used space; Devbox Plus $250/mo incl. 50,000 Devbox Minutes + 400 GB-month. Namespace Instances (general compute, unit minutes): Linux 2 vCPU/4 GB $0.002/min prepaid in a plan, $0.003/min overage; Team $100/mo (100,000 unit minutes), Business $250/mo | $0 compute when shut down; storage $0.20/GB-month (used, not provisioned) | Devbox shuts down automatically after 15 min unused (timeout adjustable)Devbox shuts down automatically after 15 min unused (timeout adjustable); paused/stopped Devbox keeps files and session metadata, running processes stop; connecting to a stopped Devbox starts it; ephemeral Devboxes also available | VM | Mar 2026 | Linux, macOS (Apple silicon) and Windows Devboxes; SSH, VS Code, Cursor, JetBrains, ZedLinux, macOS (Apple silicon) and Windows Devboxes; SSH, VS Code, Cursor, JetBrains, Zed; Amp, Claude Code, Codex, Copilot, Cursor agent and OpenCode can run inside; Computer Use for Devboxes announced 2026-11-11; MCP server and Devbox agent skill; TypeScript and Go SDK; US and Central Europe; instances up to 64 vCPU / 512 GB; 30-day free trial; Developer plan limited to 1 Devbox per user, 3 per workspace |
| Dev environment | Dev environment | Per user (seat) on an annual contract; Starter free for up to 5 users (self-hosted)Per user (seat) on an annual contract; Starter free for up to 5 users (self-hosted); Scale and Enterprise contact sales; unlimited environments and agents, never billed per environment or agent; you pay for your own cluster | $0 Okteto charge (per-seat pricing); sleeping environments scaled to zero on$0 Okteto charge (per-seat pricing); sleeping environments scaled to zero on your own cluster | Garbage Collector scales idle environments to zero and deletes unused ones: namespacesGarbage Collector scales idle environments to zero and deletes unused ones: namespaces sleep after 3 h idle and are deleted 14 days later; preview environments sleep after 1 h and are deleted after 7 days (admin-configurable) | — | Jan 2019 | On-demand Kubernetes environments for developers, coding agents and pull requests from one okteto.yamlOn-demand Kubernetes environments for developers, coding agents and pull requests from one okteto.yaml; built-in agent orchestrator runs Claude Code, Codex or custom agents in your cluster; Okteto plugin/skill for Claude Code (Agentic Workflows docs); self-hosted on EKS, GKE, AKS, Nutanix NKP, bare metal, including air-gapped; BYOC in AWS or GCP for Enterprise; AWS and Google Cloud Marketplace |
Ona (formerly Gitpod)Ona Cloud shut down 2026-10-31 Dev environment | Dev environment | Enterprise only, through sales (ona.com/pricing redirects to the contact-sales form)Enterprise only, through sales (ona.com/pricing redirects to the contact-sales form); Ona Cloud and Core shut down on 2026-10-31; environments run on runners in your own AWS or GCP account, so compute is billed by your cloud. Default AWS classes: Small m6i.large/45 GB, Regular m6i.xlarge/80 GB, Large m6i.2xlarge/100 GB, Extra Large m6i.8xlarge/200 GB, GPU Large g5.4xlarge/300 GB | — | Environments auto-stop after an inactivity timeout (default 30 minutesEnvironments auto-stop after an inactivity timeout (default 30 minutes; choices 30 min, 1 h, 3 h, 8 h or never, capped by an organization policy); archive and auto-delete follow separate policies | VM | Sep 2025 | Platform for cloud dev environments and background coding agents ("Part of OpenAI" in the site header)Platform for cloud dev environments and background coding agents ("Part of OpenAI" in the site header); environments from a Dev Container config, opened in a browser IDE, VS Code or JetBrains; Codex Agent plans, codes, tests and opens PRs; automations; prebuilds and warm pools; built-in web browser for the agent; SDK and CLI; Ona Cloud closed to new sign-ups and shut down 2026-10-31, customers pointed to Codex Cloud |
| Dev environment | Dev environment | Free $0 forever on a Kubernetes cluster you run (3 users, unlimited AI agentsFree $0 forever on a Kubernetes cluster you run (3 users, unlimited AI agents, 1,000 deployment minutes/month, preview environments); Starter $299/mo (3 users, 3,000 min); Growth $899/mo (10 users, 2 clusters, 8,000 min); Scale $2,999/mo (up to 25 users, 3 clusters, 30,000 min, $399/mo per extra cluster); paid plans billed yearly; Enterprise custom; compute is billed by your own cloud, and Qovery doesn't price per CPU core or node | Compute runs and is billed in your own cloud account | Preview environments per pull request are redeployed on every commit and deleted when thePreview environments per pull request are redeployed on every commit and deleted when the PR is merged or closed; Agent Tasks run one-off or on a schedule or webhook; auto-sleep listed among the FinOps guardrails | — | Sep 2026 | Kubernetes control plane for humans and AI agents on your own cloud (EKS, GKE, AKS, k3s, on-prem)Kubernetes control plane for humans and AI agents on your own cloud (EKS, GKE, AKS, k3s, on-prem); full-stack ephemeral and preview environments per PR with databases, seeded data and TLS; Agent Tasks (beta since 2026-09-10) run one-time or scheduled agents inside your infrastructure with network guardrails and central MCP access; MCP server, CLI, Terraform provider and 11 agent skills for Claude Code, Cursor, Codex, OpenCode and others; per-agent policy API tokens, RBAC and audit trail; Qovery calls itself 'the layer after the sandbox' rather than a code sandbox |
| Dev environment | Dev environment | $0.25 | — | Ephemeral environments are destroyed after 3 days by default (Extend resets itEphemeral environments are destroyed after 3 days by default (Extend resets it, up to 7 days; Premium plans set the number of days or turn off destroy-after-no-activity); environment pausing is an Enterprise feature; AI Sandbox sessions go pending, running, suspended, terminating, terminated; snapshots keep the filesystem and running processes and can be restored | — | — | Ephemeral full-stack preview environments for every pull request and every AI agent ('An Environment for Every AIEphemeral full-stack preview environments for every pull request and every AI agent ('An Environment for Every AI Agent'); Release MCP server lets agents create, deploy and inspect environments; AI Sandbox API (sessions, executions, snapshots, SSE events, vector memory stores, file uploads) on the environment backend; Instant Datasets; BYOC (Release creates and manages Kubernetes clusters in your own cloud account); GPUs and Release.ai templates; docker-compose onboarding; OpenAPI spec at api.release.com/openapi.json |
| Dev environment | Dev environment | Fixed monthly resource packages shared by all apps on the account: Starter $34/moFixed monthly resource packages shared by all apps on the account: Starter $34/mo regular, $7/mo new-user first purchase (2 vCPU, 2Gi RAM, 10Gi disk, 10GB traffic); Hobby $70/mo regular, $25/mo new-user (4 vCPU, 4Gi RAM, 20Gi disk, 50GB traffic); Standard $128/mo (8 vCPU, 16Gi, 50Gi disk); Pro $512/mo (16 vCPU, 32Gi); Team $2,030/mo; Enterprise $12,451/mo; 7-day free trial (4 vCPU, 4GB, 5GB storage). No 2 vCPU / 4 GiB package. | Flat monthly package | Fixed monthly resource packages; product page says "pay only when coding" but noFixed monthly resource packages; product page says "pay only when coding" but no idle/stop behavior is documented | Container | Sep 2024 | Cloud development environment; SSH from VS Code, Cursor, JetBrains Gateway or browser terminalCloud development environment; SSH from VS Code, Cursor, JetBrains Gateway or browser terminal; release as OCI image and one-click deploy to Sealos Cloud; public preview URL; snapshots ("instant restore from snapshots"); 20+ framework templates; Sealos platform open source (github.com/labring/sealos, 18.4k stars, license not SPDX-detected); environment ready "in 60 seconds" |
| Dev environment | Dev environment | Startup $500/month for 3 Drydocks (concurrent environments), extra Drydock $100/monthStartup $500/month for 3 Drydocks (concurrent environments), extra Drydock $100/month; Business and Enterprise custom (yearly); 30-day free trial | Environments scale to zero when idle; plan price is per concurrent environment | Environments scale to zero when idle and are deleted when the pull request merges orEnvironments scale to zero when idle and are deleted when the pull request merges or closes; Startup plan lists '2hr Since Last Visit (SLV)', configurable on Business | Container | Jul 2021 | Full-stack ephemeral environment per pull request, hydrated with a copy of your dataFull-stack ephemeral environment per pull request, hydrated with a copy of your data; agents manage environments through the Shipyard MCP server and CLI and test with the Playwright MCP; GitOps push-to-deploy; SSO-protected links; API, GitHub Action and CircleCI orb on Business; streaming logs and port forwarding; hosted in ISO 27001, FedRAMP, SOC 1 and HIPAA-compliant data centers; docs now at docs.stackgen.com/shipyard |
| Dev environment | Dev environment | Billed per sandbox created, per test invocation and per concurrent devboxBilled per sandbox created, per test invocation and per concurrent devbox, not by compute: Starter free (50 sandboxes/mo, 100 test invocations/mo, 10 concurrent devboxes, 1 cluster); Business $250/mo (100 sandboxes, 200 test invocations, 15 devboxes, 3 clusters; extra blocks of 50 sandboxes, 5 devboxes or 100 invocations up to $2,050/mo; SSO +$250/mo; 30-day trial); Enterprise custom; compute runs on your own cluster | No Signadot charge after creation (billed once per sandbox created)No Signadot charge after creation (billed once per sandbox created); compute is on your own cluster | Optional TTL (offset from createdAt or updatedAt) deletes the sandbox when it expiresOptional TTL (offset from createdAt or updatedAt) deletes the sandbox when it expires; without a TTL a sandbox lives until deleted; no idle scale-down described | — | Feb 2022 | Ephemeral environments for every coding-agent task and pull requestEphemeral environments for every coding-agent task and pull request; operator installs in your Kubernetes cluster and the hosted control plane sees only resource metadata; Smart Tests; Jobs run your Playwright, Cypress, Postman or k6 suites in-cluster against a sandbox; Plans (reusable validation workflows for agents); MCP server and agent skills; works with Claude Code and Cursor; local connect through devboxes; preview URLs |
| Dev environment | Dev environment | Billed per second while awake: small 1 vCPU/4 GB $0.045/hrBilled per second while awake: small 1 vCPU/4 GB $0.045/hr; standard 2 vCPU/8 GB $0.08/hr; large 4 vCPU/16 GB $0.15/hr; xlarge 8 vCPU/32 GB $0.29/hr; on your own AWS $0.01/$0.015/$0.025/$0.05 per hr plus your AWS bill; Hobby $0 with $5 credit (up to 3 small/standard VMs); Pro $20/mo incl. $20 usage (10 VMs, all sizes); Team $100/mo incl. $100 usage (unlimited, BYO AWS); Enterprise custom. No 2 vCPU/4 GB size | Disk only while asleep ('a few cents a day', no rate published)Disk only while asleep ('a few cents a day', no rate published); box removed after 16 h asleep | Box sleeps after 10 idle minutes (2 h if it runs a dev server)Box sleeps after 10 idle minutes (2 h if it runs a dev server); sleep is hibernation (memory to disk, processes resume on wake in a few seconds); removed after 16 hours asleep; Pro and Team allow always-on VMs | VM | Oct 2026 | New machine in ~3 s at any size from a repo snapshot; Ubuntu 24.04 with Node 22/24, Bun, Docker + Compose, sudoNew machine in ~3 s at any size from a repo snapshot; Ubuntu 24.04 with Node 22/24, Bun, Docker + Compose, sudo; 40 GB disk; one machine per git worktree; ports forwarded to localhost; preview links for PRs; tesser CLI plus agent skill for Claude Code and Codex; cloud agents (Devin, Capy) use an API token; bring your own compute in your AWS account/VPC; warm pools; no session time limit |
| GPU | GPU | No sandbox fee; you pay the GPU instance's normal hourly price, billed per minuteNo sandbox fee; you pay the GPU instance's normal hourly price, billed per minute, from a live marketplace (pricing page 'from $0.78/hr'; on 2026-11-14 H100 80 GB $2.98/hr on-demand, $2.20 spot; B300 $11.52/hr; H200 $5.52/hr); spend limit per AI connection | Stopped instance bills storage only (rate not given on the sandbox pages) | VM runs until you Stop or terminate it; a stopped instance bills only storage and keepsVM runs until you Stop or terminate it; a stopped instance bills only storage and keeps packages, models and checkpoints; terminate removes everything; single commands run 10 min by default, up to 24 h in the background; minimum runtime before termination applies on some offers | VM | Oct 2026 | GPU-centred: H100, H200, B200, B300, A100, RTX cards, single or multi-GPUGPU-centred: H100, H200, B200, B300, A100, RTX cards, single or multi-GPU; MCP connector for Claude, Cursor, Codex, VS Code (connect.spheron.ai/mcp); Python (spheron-sandbox) and TypeScript (@spheron/sandbox) SDKs added 2026-11-13; deploy preview and user approval before any spend; user's own SSH key works throughout; VM opens no inbound port; every command logged; boots in a few minutes |
| GPU | GPU | $0.0454 | $0 (a stopped machine is billed nothing; snapshot retention cost not published) | Billed per running second; stop snapshots the machine on the way down and stops billingBilled per running second; stop snapshots the machine on the way down and stops billing, start brings it back; optional idle auto-stop (e.g. auto_stop_seconds=900) with a keep_active heartbeat; the 50 GiB /workspace disk is the only place edits survive a stop; machines with a desktop resume cold; snapshots and forks; rm destroys the disk | VM | Sep 2026 | Machines with optional Nvidia GPUs (one card per machine with checkpointing) for coding agentsMachines with optional Nvidia GPUs (one card per machine with checkpointing) for coding agents, kernel work and GPU experiments; single-binary CLI plus Python (tama-sdk) and TypeScript SDKs; SSH; Claude and Codex in the base image, started with tama prompt; exec, logs and file APIs; HTTPS port exposure; optional browser desktop and terminal; custom images from a Dockerfile; templates; workspace secrets never stored in snapshots or forks |
| GPU | GPU | $0.166 | $0 (sandbox ends at TTL and its disk is wiped; no paused state) | Short-lived: runs until you stop it or its TTL ends (timeout set at create, e.g. 900 s)Short-lived: runs until you stop it or its TTL ends (timeout set at create, e.g. 900 s); disk wiped when it ends; scales to zero; never preempted. (Persistent Instances are a separate product with snapshots) | VM | — | GPU-centred sandboxes for RL environments, evals and GPU CIGPU-centred sandboxes for RL environments, evals and GPU CI; Python SDK (thunder_sandbox), MCP server for Claude Code, Cursor, Codex, REST API and Harbor integration; block_network option; systemd and docker compose work inside; SDK example 4 vCPU / 32 GiB / 50 GB + H100 |
705 of 705 companies