# Sandbox

Run a coding agent in a Vercel Sandbox that is already pointed at Boundless.


A coding agent needs a model and an isolated machine to run commands on. Boundless supplies the model. Vercel Sandbox supplies the machine. `@boundless.network/sandbox` wires them together so the sandbox boots with Claude Code, Codex and OpenCode already talking our gateway.

## Use it

```bash
npm i @boundless.network/sandbox @vercel/sandbox
```

```typescript
import { sandbox } from "@boundless.network/sandbox";

const box = await sandbox();
console.log(await box.run("claude -p 'add tests for src/auth'"));
await box.stop();
```

## What is in the image

| Harness | Config | Default model |
| --- | --- | --- |
| Claude Code | `~/.claude/settings.json` | `glm-5.2` |
| Codex | `~/.codex/config.toml` | `qwen3.6` |
| OpenCode | `~/.config/opencode/opencode.json` | `glm-5.2` |

- `Ubuntu 26.04`
- `Node 24`
- `Python 3`
- `git`, `ripgrep`, `jq`, and `curl`

`OPENAI_BASE_URL` and `ANTHROPIC_BASE_URL` already point at the gateway, so the OpenAI and Anthropic SDKs work inside the sandbox with no configuration either.

## Default budget

Minted keys are capped at $5 and expire after an hour; they appear in [API Keys](/api-keys). Sandbox compute is billed by Vercel.