vercel-sandbox（Vercel sandbox 搜索）

Vercel Sandbox Skill

Vercel Sandbox is an ephemeral compute primitive (Beta, all plans) that runs untrusted or user-generated code in isolated Linux microVMs on Vercel. It supports AI agents, code generation tools, developer sandboxes, and backend logic testing.

When to consult reference files

• Getting started / setup: Read references/quickstart.md
• SDK usage (TypeScript): Read references/sdk-reference.md
• CLI commands: Read references/cli-reference.md
• Pricing, limits, resource specs: Read references/pricing-and-specs.md
• Authentication details: Covered in both quickstart and SDK reference files

Quick reference

Install

pnpm i @vercel/sandbox    # TypeScript
pip install vercel-sandbox # Python

Authentication setup

vercel link         # Link to a Vercel project
vercel env pull     # Pull OIDC token to .env.local (expires 12h locally)

Minimal example (TypeScript)

import { Sandbox } from '@vercel/sandbox';

const sandbox = await Sandbox.create();
const result = await sandbox.runCommand('echo', ['Hello from Vercel Sandbox!']);
console.log(await result.stdout());
await sandbox.stop();

Key SDK classes

Available runtimes

node24, node22, python3.13 — all on Amazon Linux 2023.

Default working directory

/vercel/sandbox — user code runs as vercel-sandbox user.

Timeouts

Default 5 min. Max: 45 min (Hobby), 5 hours (Pro/Enterprise). Extend with sandbox.extendTimeout().

Snapshots

Capture with sandbox.snapshot() — sandbox stops after snapshot. Expire after 7 days.

Updating documentation

When the user asks to "update the Vercel Sandbox documentation", choose the right method based on the environment:

Method selection

• claude.ai / Claude Desktop (has web_fetch tool) → Use Path A: web_fetch
• Claude Code (has bash with internet) → Use Path B: Python script

To detect: if web_fetch is in your available tools, use Path A. Otherwise use Path B.

Path A: web_fetch (claude.ai / Claude Desktop)

1. Use web_fetch to retrieve https://vercel.com/docs/vercel-sandbox (the main page)
2. Discover all sub-page links under /docs/vercel-sandbox/
3. Fetch each known sub-page URL (see list below) plus any newly discovered pages
4. For each reference file in references/: a. Read the existing file content b. Compare with fetched content — note additions, removals, changes c. Rewrite the reference file with updated, clean markdown d. Update the > Last fetched: date at the top
5. Report to user: pages fetched, what changed per file, update date

Path B: Python script (Claude Code / local terminal)

Run the update script which has full internet access:

python3 <skill-path>/scripts/update_docs.py <skill-path>/references

The script will:

1. Fetch all known sandbox doc URLs from vercel.com
2. Discover any new sub-pages from the main page
3. For each reference file: compute a diff, overwrite with new content, report changes
4. Print a summary of what changed and when

After the script runs, review its output and report changes to the user.

Known doc page URLs

• https://vercel.com/docs/vercel-sandbox (overview + system specs)
• https://vercel.com/docs/vercel-sandbox/quickstart
• https://vercel.com/docs/vercel-sandbox/sdk-reference
• https://vercel.com/docs/vercel-sandbox/cli-reference
• https://vercel.com/docs/vercel-sandbox/pricing
• https://vercel.com/docs/vercel-sandbox/examples
• https://vercel.com/docs/vercel-sandbox/managing

Also discover any new pages linked from the main page's sidebar/navigation.