Public beta: free for registered users

Sandboxes for AI agents

Give your agent a real Linux machine for every task. edshell starts a Firecracker microVM in under a second, lets the agent run commands and move files in it, then throws it away.

Get an API key Read the API reference

Or from your terminal: npx edshell login 

import { Sandbox } from 'edshell-sdk';

// EDSHELL_API_KEY from the environment
const sb = await Sandbox.create({ template: 'python' });

const { exitCode, stdout } = await sb.run('python -c "print(6*7)"');
// 0 '42\n'

await sb.files.write('/tmp/data.json', json);
for await (const ev of sb.exec('pip install pandas')) {
  process.stdout.write(ev.text);
}

await sb.kill();
Warm start
< 1 second
Isolation
Own kernel per sandbox
Billing unit
Per second
Price during beta
$0

A computer your agent can break

Code written by a model should not run on your servers. Each edshell sandbox is a separate virtual machine, made for one task and gone when the task ends.

MicroVMs, not containers

Firecracker, the VMM behind AWS Lambda. Every sandbox has its own kernel and runs inside a jailer, so a breakout has to escape a hypervisor.

Fast from a warm pool

Booted VMs wait in a pool, so create usually returns in under a second. Docker sandboxes start fresh in about 20 s.

Run, stream, read and write

Run commands to completion or stream their output, read and write files, or open an interactive terminal over a WebSocket.

Root and the network

Passwordless sudo, outbound internet, and package managers that work: apt, pip, npm, docker.

Gone when you are done

Kill it yourself, or let its timeout do it, so a sandbox you forget ends on its own. Nothing is kept after a sandbox ends.

A short path for data

The API only creates and ends sandboxes. Commands and files go straight to the host the sandbox runs on, with a token scoped to that sandbox.

How it works

  1. 01

    Get a key

    Run npx edshell login and approve the code, or sign in with Google and create a key in the dashboard (EDSHELL_API_KEY).

  2. 02

    Create a sandbox

    Choose a template, and optionally the vCPUs, memory and a timeout. You get back the sandbox id and a token for its host.

  3. 03

    Use it, then end it

    Your agent runs commands and moves files. Kill the sandbox when the task is done; usage stops being counted that second.

Templates

All are Ubuntu 24.04 with a user that has sudo. Choose a size per sandbox: up to 2 vCPUs and 2 GiB during the beta.

Template What is inside Default size
base Ubuntu 24.04, git, curl, build tools 1 vCPU, 512 MiB
python base + Python 3.12, pip, venv 1 vCPU, 1 GiB
node base + Node.js 22 LTS, npm, pnpm 1 vCPU, 1 GiB
docker base + Docker Engine, Compose, buildx 1 vCPU, 1 GiB

Free while we are in beta

Sign in and use it. We record usage per second (vCPU-seconds and GiB-seconds) and show it in your dashboard, so you will know what it would cost before paid plans start. Nothing is billed during the beta.

Beta

$0 / month

  • ✓2 sandboxes running at once
  • ✓Up to 2 vCPUs and 2 GiB of memory each
  • ✓Up to 1 hour per sandbox
  • ✓5 API keys
Get started