Chromium Bridge is a two-part software package: a Node.js MCP (Model Context Protocol) server and a Manifest V3 Chromium extension. Together they allow Claude Code (Anthropic's CLI coding agent) to control a locally running Chromium-based browser — listing tabs, navigating pages, reading text and accessibility trees, taking screenshots, clicking, typing, filling forms, running JavaScript, capturing console/network logs, and even recording GIFs. It also ships a chat panel popup so the user can talk to Claude directly from the browser toolbar.

The problem being solved is a compatibility gap. Anthropic's official 'Claude in Chrome' extension relies on the Chrome tab groups API, which is missing or broken in forks like Arc, Vivaldi, and Brave. This project reimplements the same browser-automation surface using only the plain chrome.tabs, chrome.scripting, and chrome.debugger APIs, so it works in any Chromium browser that can load an unpacked extension.

Users are developers who already run Claude Code and want agentic browser automation in a non-standard Chromium browser. Installation is straightforward: load the extension unpacked into the target browser profile, then register the MCP server with `claude mcp add ... npx chromium-bridge`. Once connected, the browser_* tools become available to Claude, and clicking the extension icon opens a chat panel backed by the Claude Agent SDK.

Architecturally, Claude Code speaks stdio MCP to server/index.mjs. That server exposes browser_* tools and forwards commands over a WebSocket on 127.0.0.1:8929 to the extension's service worker. The service worker executes commands via chrome.debugger (CDP) — enabling real mouse events, off-screen screenshots, and network/console capture — plus chrome.tabs and chrome.scripting for basic tab management. A second WebSocket path (/chat) drives the popup chat UI through the Claude Agent SDK, reusing the same tool set with an optional 'Ask before acting' permission prompt.

Notable engineering details: the WS server rejects any Origin that is not chrome-extension://, keeping arbitrary web pages out (though the README honestly notes it cannot distinguish extensions or block other local processes). Visual on-page indication (orange glow, virtual cursor) is injected during agent actions and hidden from screenshots. The project is honest about its trust model and single-profile / single-session limitations. No red flags observed — code matches the stated purpose.