PawBrowse

PawBrowse connects your local AI coding agent — such as Claude Code — to your real, logged-in Chrome so it can read and act on web pages the way you would. Instead of screenshots, PawBrowse reads each page as a compact, numbered element table (e12 "Sign in", e7 "Email", ...), so your agent can click, type, select, navigate and verify quickly and reliably. The agent is the policy. PawBrowse does not run its own AI model and requires no API key. It talks only to a local bridge on your own machine (127.0.0.1). Page content is passed to the agent you are running and is never sent to the developer or any third-party server. WHAT YOU NEED - The open-source PawBrowse MCP server running locally (Node.js). See the project page. - An MCP client such as Claude Code. HOW IT WORKS 1. Run the PawBrowse MCP server (your AI client launches it). 2. This extension connects to it over a local WebSocket (127.0.0.1). 3. Your agent calls tools: observe (element table), act (click/type/select/scroll), navigate, and assert. FEATURES - Drives your real, logged-in browser via the Chrome DevTools Protocol — no separate debug port and no browser relaunch. - Element-table perception with stable refs (numbered controls, not screenshots). - No external model, no API keys, no telemetry. - Fully open source (MIT). Source, docs and issues: https://github.com/ItaiZeilig/pawbrowse Privacy policy: https://github.com/ItaiZeilig/pawbrowse/blob/main/PRIVACY.md PawBrowse is an independent open-source project and is not affiliated with or endorsed by Google or Anthropic.