Give your AI agents secure microVMs to run code and control desktop apps. Each sandbox boots in under a second with its own kernel — not a container, not a shared runtime.
$ sandbox.create()
Booting microVM...
✓ Ready in 247ms
$ sandbox.commands.run("python3 agent.py")
Running in isolated kernel...
{"status": "complete", "output": "..."}
$ sandbox.kill()
✓ Destroyed. No state remains.
$ _
Powering code playgrounds for
Built for performance
0ms
VM restore
0
Playgrounds
0h
Free tier / month
0
Shared state
Developer experience
$ npm install @omnirun/sdk
$ cat run.ts
import { Sandbox } from '@omnirun/sdk'
const sandbox = await Sandbox.create('playground')
const result = await sandbox.commands.run('echo "Hello from a microVM"' )
console.log(result.stdout)
await sandbox.kill()
$ npx tsx run.ts
Hello from a microVM
✓ Sandbox killedHow isolation works
01
One API call. Typical boot under a second. Full Linux environment with its own kernel.
02
Execute code, install packages, read and write files. Nothing shared between sandboxes.
03
Destroyed completely. No state leaks. No cleanup. Gone.
Use cases
Capabilities
Each sandbox runs its own Linux kernel inside a microVM, the same technology behind AWS Lambda. Hardware-level isolation, not process boundaries.
Untrusted code gets no network by default; agent templates get full egress. Lock either down to a per-domain allow-list with the opt-in SNI proxy.
Upload and download URLs expire. Access tokens scoped to a single sandbox.
Claude Managed Agents
Anthropic runs the agent loop and model. Every tool call executes in an isolated microVM on your own infrastructure — its own kernel, your perimeter, EU-resident, with optional per-domain egress control.
Free tier includes 25 sandbox-hours per month. No credit card required.
Get Free API Key