browser-pilot

Give AI agents a real browser. One command to install.

A lightweight MCP server that wraps chrome-devtools-mcp with automatic Chrome lifecycle management — works with Claude Code, Codex CLI, and Gemini CLI out of the box.

What can you do with it?

• Deep web research — crawl multi-page sites, extract structured data, summarize articles, compare prices across tabs
• Form automation — fill and submit forms, upload files, interact with SPAs and dynamic content
• Human-in-the-loop — AI handles the tedious parts, pauses when a CAPTCHA or login appears, you solve it, AI continues
• Screenshot & visual analysis — capture pages and let the AI reason about layout, content, or visual changes
• Scraping without APIs — access any site a human can open, no API key required

Requirements

• Node.js ^20.19.0, ^22.12.0, or >=23 (required by chrome-devtools-mcp)
• npm >=7

Install

npx --package=@yqi96/browser-pilot@latest browser-pilot-install

Detects which AI clients you have installed and registers itself automatically.

# Target a specific client
npx --package=@yqi96/browser-pilot@latest browser-pilot-install --client claude
npx --package=@yqi96/browser-pilot@latest browser-pilot-install --client codex
npx --package=@yqi96/browser-pilot@latest browser-pilot-install --client gemini

What it does

• Auto-launches Chrome on first use (macOS / Linux / Windows)
• Adds lifecycle tools — browser_open, browser_close
• Proxies all chrome-devtools-mcp tools transparently (navigate, click, screenshot, fill forms, …)
• Returns screenshots as base64 for direct multimodal AI processing
• Parallel-agent safe — each agent gets its own isolated Chrome instance via session IDs

Parallel agents

When multiple subagents run concurrently, each one calls browser_open() independently and gets back a unique session_id. Every subsequent browser tool call passes that ID via _browser_session, so agents never interfere with each other:

# Agent A                                    # Agent B (concurrent)
browser_open()                               browser_open()
→ Session: abc-123                           → Session: def-456

navigate_page(url="...",                     navigate_page(url="...",
  _browser_session="abc-123")                  _browser_session="def-456")

browser_close(session_id="abc-123")          browser_close(session_id="def-456")

If Chrome is closed externally mid-session, the next tool call returns a clear error telling the agent to call browser_open() again — no manual cleanup needed.

Supported clients

| Client | Status | Skill command | |--------|--------|---------------| | Claude Code | ✅ Supported | /browser | | Codex CLI | ✅ Supported | $browser | | Gemini CLI | ✅ Supported | activate_skill("browser") |

Uninstall

npx --package=@yqi96/browser-pilot@latest browser-pilot-uninstall

Contributing

Found a bug? Open an issue or, if you have Claude Code, clone the repo and let it fix it:

git clone https://github.com/yqi96/browser-pilot && cd browser-pilot && claude

PRs are welcome!

License

MIT