The sandbox platform for AI agents.

Sandchest gives your agent isolated Linux VMs that fork in under 100ms. It tries ideas in parallel, backtracks bad paths instantly, and iterates faster than any human could.

bun add @sandchest/sdk

How it works

1
Add Sandchest to your agent — install the SDK or connect the MCP server. Your agent gets sandbox tools automatically.
2
Your agent iterates faster — it spins up VMs, runs code, and forks state in milliseconds. No waiting, no setup. Faster feedback loops mean more iterations per task.
3
Your agent gets smarter — it explores multiple approaches in parallel, discards bad paths instantly, and keeps the best result. More attempts, better outcomes.

What is Sandchest?

A sandbox platform for AI agent code execution. Every sandbox is a Firecracker microVM with VM-grade isolation, sub-second fork capability, and a permanent session replay URL.

Sub-100ms forking

Fork a running sandbox — memory, disk, everything — in under 100ms.

VM-grade isolation

Every sandbox is a real Firecracker microVM. Full hardware-level isolation.

TypeScript SDK

Create, exec, fork, and manage sandboxes with a few lines of code.

Session replay

Every session is fully replayable. Logs, file changes, and terminal output.

MCP server

First-class MCP support for Claude Code and other AI tools.

CLI

Create, exec, fork, and SSH into sandboxes from your terminal.

See it in action

A few lines of TypeScript. That's it.

agent.ts

import Sandchest from "@sandchest/sdk";

const sandchest = new Sandchest();

// Create a sandbox and set up the environment
const sb = await sandchest.create();
await sb.exec("git clone repo && npm install");

// Try something risky in a fork
const fork = await sb.fork();
const result = await fork.exec("npm test");

if (result.exitCode !== 0) {
  await fork.destroy(); // original untouched
}

Follow along

Sandchest is being built in the open. Star the repo to follow development and get notified when we ship.

Star on GitHub View Repository