Browser Driver MCP

Browser Driver MCP connects your real, signed-in browser to an AI agent running on your own computer — Claude Code, Cursor, or any Model Context Protocol (MCP) client — so it can read pages and get things done in your actual tabs. It offers the same tools as Claude in Chrome, with the same names, parameters and behavior, so prompts and skills written for Claude in Chrome work unchanged: screenshots, clicking, typing, keyboard shortcuts, scrolling, drag and drop, reading the page as an accessibility outline, finding elements by description, filling forms, uploading files, running JavaScript, and reading console and network activity. It can also record annotated GIFs of what it does. YOU STAY IN CONTROL • The agent only works in a dedicated "MCP" tab group — your other tabs are never touched. • Every page it drives glows electric blue and shows a Stop button. One click detaches it immediately. • Password, payment and one-time-code fields are redacted from what the agent reads. PRIVATE AND LOCAL • No servers, no accounts, no analytics. The extension talks only to a small helper program on your own computer, through Chrome's native messaging. • It opens no network ports, so web pages and other computers can't reach it. WORKS WHERE OTHERS DON'T • Reads inside shadow DOM (including closed shadow roots) and cross-origin iframes. • Handles alert/confirm dialogs so automation never gets stuck. SETUP After installing, a setup page shows live status and the one line to add to your MCP client, e.g. for Claude Code: claude mcp add browser-driver -- npx -y browser-driver-mcp The server sets up its connector to this extension automatically the first time it runs (requires Node.js 18+). Full instructions: https://github.com/darrinm/browser-driver-mcp Open source (MIT). Independent project; not affiliated with Google or Anthropic.