Introduction

What Sandal is#

Sandal is the execution layer for AI agents. Every time your code (or your agent) needs to run something, a shell command, a test suite, a mobile app, an untrusted snippet of Python from an LLM, it needs somewhere safe to run it. That's Sandal.

Under the hood, every sandbox is its own Firecracker or Cloud Hypervisor microVM with its own kernel. No shared containers, no noisy neighbors. We run on bare metal (Hetzner, OVH, Leaseweb, and dedicated Mac minis for the Apple stuff) instead of renting margin from a hyperscaler, which is why the pricing looks the way it does: $0.003/min for a 2vCPU Linux sandbox or runner, paused sandboxes cost $0.

One API key. One SDK. Four surfaces.

The four surfaces#

Every surface is the same primitive underneath, a microVM you create, use, and either destroy or pause. Paused instances don't bill compute; you only pay again when you claim them back.

60-second quickstart#

Install the CLI and scaffold a project:

$ npx sandal init

This drops an API key prompt in your terminal (or reads SANDAL_API_KEY if it's already set) and writes a starter script for the SDK you pick.

npm install @sandal/sdk

import { Sandal } from "@sandal/sdk";

const sandal = new Sandal({ apiKey: process.env.SANDAL_API_KEY });

const sb = await sandal.sandboxes.create({ template: "node-24" });
const result = await sb.exec("pnpm test");
console.log(result.stdout);

const url = await sb.expose(3000);
console.log(`Preview: ${url}`);

await sb.pause(); // billing stops now

That's a real sandbox, running real code, in under a second of cold-start latency, see the quickstart for the full walkthrough including forking and API key setup.

Where to go next#

• Quickstart, install, authenticate, create your first sandbox
• Sandboxes, the core primitive: exec, fs, expose, pause, fork
• Runners, drop-in GitHub Actions runners on bare metal
• Simulators, iOS/Android devices you can control from code
• Architecture, how the microVMs, snapshots, and metering actually work
• Pricing, evidence-gated numbers, every figure from a qualification run

Every page on these docs is also available as raw markdown by appending .md to the URL, and the whole set is summarized at /llms.txt for feeding into your own agent.