Use cases

What you can build when every agent has a computer of its own.

A spawn sandbox is a Firecracker microVM your agent can’t break out of: its own kernel, memory and disk, created in one call and wiped when you’re done. Here’s what that makes possible.

01

Agent swarms on your repo

Connect a GitHub repo and hand off a goal. An orchestrator splits it into tasks, every agent works in its own sandbox, and the work comes back as one pull request.

  • One sandbox per task. Create one for each agent and destroy it when its task lands. Nothing leaks from one task into the next.

  • A real machine to work in. Agents run git, package managers and your test suite in a persistent shell, with a filesystem of their own.

  • Step in whenever you want. Open a live terminal, or VS Code in the browser, on any agent’s sandbox while it works.

Create a sandbox
02

Security review on every PR

When a pull request opens, an agent checks it out in a fresh sandbox, runs the tests, hunts for vulnerabilities and comments on the exact line, before a human looks.

  • A clean machine for every review. Each review starts in a fresh sandbox that’s destroyed afterwards, disk and all, so one PR can’t affect the next.

  • Untrusted code stays contained. The PR’s code runs inside its own microVM. Switch network access off when it shouldn’t reach the internet.

  • Real tools, real findings. Run tests and scanners with exec, stream long jobs over a terminal, and post the results back to GitHub.

Run code and commands
03

RL training environments

Run every episode in its own sandbox, restored from a clean snapshot. A policy that learns to game its reward can’t reach the host or the other environments.

  • The same start, every episode. Each sandbox boots from a prebuilt snapshot, so every rollout begins in exactly the same environment.

  • Isolation a policy can’t game. Every environment is a separate microVM with its own kernel, so a misbehaving policy stays inside its own cell.

  • Pause long episodes. Pause an environment to free its CPU and memory, then resume it later on the same machine with its disk intact.

Pause, resume, destroy
04

From your terminal or your code

Script sandboxes from the spawn CLI today, or drive them from your own agent with the Python and TypeScript SDKs, coming soon. Same API, same isolation.

  • The CLI. Create, exec, pause, resume and destroy sandboxes from your shell or CI, with JSON output for scripts.

  • Plain HTTP underneath. Every operation is one HTTPS call, so any language can use spawn today.

  • SDKs on the way. Until the Python and TypeScript SDKs ship, the docs show every call in cURL, Python and JavaScript.

Quickstart

More ways to use a sandbox

Code interpreter Give a data agent Python or Node that stays warm between calls, so variables and loaded data carry over. Eval harness Run each benchmark task in a fresh machine, so one task’s leftovers never skew the next result. Dev environments Open VS Code in the browser inside a sandbox, with the agent’s files and terminal right there. Grading untrusted code Run student or user submissions with networking switched off, then wipe the machine. App previews When an agent starts a web server, share it as a private preview link that expires on its own. Long-running agents Agents that work for hours pause when idle and pick up where they left off, disk intact.

Give every agent somewhere safe to work.

Get early access Read the docs

Private beta · $20 in free credit · no card required