Every free agent sandbox tier that clears our bar in 2026: 4 offers meet the criteria and 6 offers are demoted with a named reason. Our membership test: isolated execution environments provisioned programmatically, sold for AI agent workloads. We could not confirm today's terms for 1 of them, and on all 1 our own read did not confirm them. Each row says why.
Free tier with $5.40/mo usage credit. No card.
Listed in Cloud Hosting, and on this page because we labelled it agent_sandbox. We read that on 2026-08-31 from https://www.pandastack.io/, where it says: “Firecracker microVMs that boot in 49ms. Agent sandboxes with fork & snapshot, git-driven app hosting, dedicated Postgres 16 and serverless functions on one open-source substrate.”. How we use this
$1 of free credit a month on Railway's Free plan, after a one-time $5 grant on the 30-day trial; the credit is shared with other Railway usage and does not roll over. Sandboxes are isolated Linux VMs for coding agents, test runs or code execution, managed by SDK, CLI, dashboard or API, with Claude Code, Codex, OpenCode and Pi preinstalled. Free and Trial plans allow 10 sandboxes per environment, each up to 2 vCPU and 2 GB. The idle timeout is at most 5 minutes, after which Railway destroys the sandbox. Rates are $50 per vCPU-month and $50 per GB-month for CPU used and memory in use, metered per second, and egress at $0.05 per GB; a 20-minute task averaging 1 GB of memory and 30% of one vCPU costs about 3 cents. Hobby costs $5 a month with $5 of usage, allows 50 sandboxes per environment up to 8 vCPU and 8 GB, and an idle timeout up to 120 minutes or off. When we last read the page we cite for this offer, on 2026-11-09, we found a change we could not reconcile with the terms we publish, so we cannot confirm these terms today.
Listed in Cloud Hosting, and on this page because we labelled it agent_sandbox. We read that on 2026-11-03 from https://docs.railway.com/sandboxes, where it says: “Sandboxes are isolated Linux VMs you create on demand, run commands in, and destroy.”. How we use this
Isolated Linux microVMs for running untrusted or agent-generated code, metered by active CPU, memory, sandbox creations and data transfer. The Hobby plan includes each month at no cost: 5 hours of Active CPU, 420 GB-hours of provisioned memory, 5,000 sandbox creations and 20 GB of data transfer, plus 15 GB of snapshot storage for the life of the account. Hobby sandboxes run for at most 45 minutes, 10 at a time. Past the quota, sandbox creation pauses and nothing is charged. On Pro, Active CPU costs $0.128 an hour and memory $0.0212 per GB-hour.
Listed in Cloud Hosting, and on this page because we labelled it agent_sandbox. We read that on 2026-09-27 from https://vercel.com/docs/sandbox, where it says: “Run untrusted or agent-generated code in isolated Linux microVMs.”. How we use this
Serverless platform for running CPU and GPU workloads. The Starter plan has no monthly fee and includes $30 of compute per month (GPU, CPU and memory), 3 workspace seats, 100 containers and 10 GPU concurrency. Usage beyond the $30 is billed per second: T4 $0.000164 per second; H100 $0.001097 per second. Since 2026-09-01, Shared Endpoints are billed per token from the first request and are not covered by the $30. Since 2026-10-01, network egress beyond 1 TiB a month is billed at $0.04 per GiB. Team is $250 per month with $100 of compute included. Note — 2026-10-01: Since October 1, 2026, Modal charges for network egress: the Starter plan includes 1 TiB a month, Team 10 TiB and Enterprise 100 TiB, and egress beyond that costs $0.04 per GiB. Source ↗
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-10-31 from https://modal.com/pricing, where it says: “Isolated environments for coding agents, RL rollouts, and untrusted code”. How we use this
Sandboxes for running AI-generated code, priced pay as you go. New accounts get $200 in free compute on a free trial that needs no credit card; Daytona does not say whether the credit expires or renews, and it does not cover GPU sandboxes. A sandbox can use up to 4 vCPUs, 8 GB of RAM and 10 GB of disk, and on the first two usage tiers its network access is restricted. After the credit: $0.0504 an hour per vCPU and $0.0162 an hour per GiB of memory; the first 5 GiB of storage is free.
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-09-27 from https://www.daytona.io/pricing, where it says: “Daytona lets you spin up sandboxes in milliseconds and shut them down just as fast.”. How we use this
$30 in free trial credit, one per user and at most one per org. Sprites are persistent, hardware-isolated Linux VMs on Fly.io for AI agents, dev environments and untrusted code, with sign-in through a Fly.io account. Compute is billed only while a Sprite runs; it pauses after about 30 seconds idle, and a paused Sprite costs nothing beyond its storage. Rates are $0.0385 per CPU-hour, $0.021875 per GB-hour of memory, $0.000683 per GB-hour of hot storage while awake and $0.000027 per GB-hour of cold storage. There is no per-Sprite charge, and bandwidth is not metered. Each Sprite has up to 8 vCPUs and 100 GB of storage. Fly.io does not say whether the credit needs a card.
Listed in Cloud Hosting, and on this page because we labelled it agent_sandbox. We read that on 2026-11-03 from https://docs.fly.io/sprites, where it says: “Sprites are persistent, hardware-isolated Linux environments for running arbitrary code.”. How we use this
Devboxes: secure micro-VM environments for building and running AI agents. New accounts get $50 in usage credits with Pro features and no credit card; usage draws the credit down until it is gone, and nothing is charged automatically when it ends. On the trial: 3 running devboxes, sizes up to Medium (2 CPUs, 4 GB), and at most 1 hour of keep-alive. After the credit, the Basic plan is a free subscription with usage billed at $0.108 per CPU-hour and $0.0252 per GB-hour of memory. Pro is $250 a month plus usage.
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-09-27 from https://runloop.ai/pricing, where it says: “Secure micro-VM environments to build and run AI agents at speed”. How we use this
Sandboxes and runtime for AI agents, priced on usage with no plan fee. New accounts get up to $200 in free credits; Blaxel does not say whether they renew, and its terms say unused service credits expire one year after they are issued unless otherwise specified. The free quota tier allows 10 sandboxes, which it deletes after at most 7 days, and free accounts cannot request higher quotas. A running sandbox costs $0.0000115 per GB of RAM per second; idle time is not billed as active CPU.
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-09-27 from https://blaxel.ai/pricing, where it says: “Up to $200 in free credits. Spin up your first sandbox in under a second.”. How we use this
Cloud sandboxes for AI agents: isolated Linux machines that run generated code and shell commands, work with files and reach the internet. The Hobby plan has no monthly fee and a one-time $100 usage credit, with no credit card needed to start. When the credit runs out, the account is blocked until a payment method is added. Hobby allows 20 concurrent sandboxes and up to 1 hour of continuous runtime per sandbox. Running sandboxes are billed per second for CPU and RAM. Pro is $150/month plus usage.
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-09-01 from https://e2b.dev/pricing, where it says: “We charge per second of a running sandbox”. How we use this
Sandboxes for AI agents from Novita AI, billed per second for CPU and RAM while a sandbox runs. New users who complete an account setup survey get $100 in promotional credits for Novita Sandbox usage, with no credit card; the credits expire, and Novita does not state after how long. The free tier allows 5 concurrent sandboxes of up to 2 vCPUs and 4 GB of RAM, each running for at most 1 hour. Paid usage costs $0.0000098 per vCPU-second and $0.0000032 per GiB-second.
Listed in AI / ML, and on this page because we labelled it agent_sandbox. We read that on 2026-09-27 from https://docs.novita.ai/guides/sandbox-pricing, where it says: “Novita Sandbox is billed per second for CPU and RAM while a sandbox is running.”. How we use this
| Vendor | Free Tier | Key Limits | Durability | Read / catalogue date |
|---|---|---|---|---|
| PandaStack | Free | Free tier with $5.40/mo usage credit. No card. | — | 2026-10-29 |
| Railway Sandboxes | Free | $1 of free credit a month on Railway's Free plan, after a one-time $5 grant on the 30-day trial; the credit is shared wi... | not published read not reconciled · 2026-11-09 |
2026-11-09 / 2026-11-03 |
| Vercel Sandbox | Hobby | Isolated Linux microVMs for running untrusted or agent-generated code, metered by active CPU, memory, sandbox creations ... | stable | 2026-11-12 |
| Modal | Free | Serverless platform for running CPU and GPU workloads. The Starter plan has no monthly fee and includes $30 of compute p... | volatile effective 2026-10-01 |
2026-11-08 / 2026-09-22 |
We publish no ranking of these. Nothing we record measures popularity or generosity, so naming a best or a top few would be our preference rather than our data. We hold 4 free agent sandbox offers that clear our bar, each listed with the terms we publish and the date we last read them. Our ranking rule is published at /criteria.
4 offers meet our criteria on this page. Of those, 1 carries no recorded change to the terms we publish, 1 a recorded narrowing, and we publish no durability signal for 2. The Durability column below says which offer is in which group, and links the dated record behind every change. We withhold the class where the pricing page does not resolve or states no amount, tier or rate we can read, where we refused the last read, where the listing is gated, or where a narrowing in our records cites no source or no archived copy of the vendor's page has confirmed it. We do not rank these by how much you get, because the limits are not comparable without reading each vendor's page.