BrowserCtrl

About

BrowserCtrl turns your real Chrome into an MCP server.
Stop spinning up sandboxed Chromiums every time your AI agent needs the web. BrowserCtrl turns the Chrome you're already using — same profile, same cookies, same tabs — into a Model Context Protocol (MCP) server that Claude Desktop, Cursor, Windsurf, Zed, and Claude Code can drive.

What BrowserCtrl does
  
Install the extension and the BrowserCtrl CLI (`npm i -g browserctrl`), and your MCP-aware AI host gets a full browser-automation surface: navigation, interaction, ARIA snapshots, network capture, screenshots, assertions, locator management — 35 tools out of the box.

Why install BrowserCtrl

You'll want this if you:

- Use Claude Desktop, Cursor, Windsurf, Zed, or Claude Code and want a "drive my browser" tool that respects your active session. 
- Build agentic workflows where the agent needs to act inside your logged-in dashboards, private CRMs, team Slack, or paid SaaS.
- Are tired of re-authenticating into every service every time the agent boots a fresh sandbox.
- Need automation that survives small UI changes — BrowserCtrl ships self-healing locators that re-anchor against ARIA snapshots when the DOM shifts.
- Care about token cost: BrowserCtrl's `--lean` mode trims the tool surface to 10 essentials and costs 2,514 host-context tokens on a control fixture versus the equivalent Playwright MCP setup at 5,446 — roughly half. Across an agent loop, the savings compound.

How it works
  
1. Install the BrowserCtrl CLI: `npm i -g browserctrl`
2. Run the one-shot wizard: `browserctrl init`. It installs the Chrome Native Messaging host, registers BrowserCtrl with your AI hosts (Claude Desktop, Cursor, etc.), and validates the connection.
3. Install this Chrome extension. It auto-connects to the local daemon via Chrome Native Messaging — no pair code, no token to copy, no localhost socket to manage.
4. Open your MCP host and ask it to do something in your browser. ("Use browserctrl to summarize my open GitHub PRs.")

The entire flow is local. There's no BrowserCtrl-operated cloud service, no auth server, no signal sent off-machine.

Security and privacy

- Per-origin consent gate — by default, every new origin the agent operates on prompts you in the popup. You can pin allow/deny per host.
- Hard refusals on destructive keyboard chords (close tab, quit Chrome, new
- Rate-limited — 16 actions/minute per tab, 60/minute globally — to prevent runaway agent loops.
  
Requirements
  
- Chrome 120 or later
- Node.js 20.10 or later (for the local CLI / daemon)
- macOS or Linux (Windows support coming)

Support
  
Closed alpha — bugs are expected. The fastest way to get a fix is to file an issue: https://github.com/DeRaowl/browserctrl/issues
  
Direct contact: [email protected]