<img src="extension/icons/icon128.png" width="32" alt="" align="top"> Chromium Bridge

Русская версия

A bridge between your Chromium-based browser and Claude Code. The official "Claude in Chrome" extension connects in some Chromium browsers (Arc, Vivaldi, and others) but automation hangs: its tools are built on the tab groups API, which is missing or broken there. This bridge uses only plain chrome.tabs / chrome.scripting / chrome.debugger, so it works in any Chromium browser that can load an extension.

Demo

Claude driving the browser through the bridge — opening Wikipedia, typing a search, and landing on the article:

Architecture

Claude Code ⇄ (stdio MCP) ⇄ server/index.mjs ⇄ (WebSocket, 127.0.0.1:8929) ⇄ extension in the browser
                                   ⇅ (WebSocket /chat)
                            chat panel (popup on the extension icon)

• extension/ — an MV3 extension: the service worker keeps a WebSocket to the local server and executes its commands (tabs, navigation, page text, screenshots, clicks, form filling). Clicking the icon opens the chat panel (chat.html) — a popup anchored to the extension icon.
• server/ — an MCP server (stdio) that exposes the browser_* tools to Claude Code and proxies them to the extension. It accepts WS connections only from a chrome-extension://… Origin — regular web pages cannot connect. It also serves the /chat channel: panel messages run through the Claude Agent SDK (authenticated via the Claude Code login) with the same browser_* tools; built-in tools (Bash, Read, etc.) are disabled.

Chat panel

An equivalent of the "Claude in Chrome" side panel: a popup that opens when you click the extension icon (no chrome.sidePanel — it is not supported everywhere). The chat can see the browser: list tabs, read pages, take screenshots, and click.

The panel UI is in English by default and switches to Russian automatically when the browser UI language is Russian. A language selector (Auto / English / Русский) in the bottom bar overrides auto-detection; the on-page badge follows the same choice.

• The popup closes when it loses focus (clicking the page) — that is browser behavior. The conversation context is not lost: the panel remembers the session_id and the server resumes the conversation via the Agent SDK resume. A turn that is in flight when the popup closes is interrupted.
• It only works while the server is running (usually an active Claude Code session with the chromium-bridge MCP); otherwise the panel shows "Server unavailable".
• Model picker in the panel header: "Default" takes the model from ~/.claude/settings.json (whatever was set via /model; SDK sessions do not read Claude Code settings themselves, the server passes the model explicitly), the other entries are hard overrides. Switching applies on the fly (setModel) and is remembered. Startup override: CHROMIUM_BRIDGE_CHAT_MODEL=sonnet in the server environment. Port: CHROMIUM_BRIDGE_PORT (8929 by default) — server side only; the extension always connects to 8929, so changing the port also means editing WS_URL in extension/sw.js and extension/chat.js.
• Chat history: the 🕓 button in the header lists past conversations (stored in the panel's localStorage, the last 30).
• After each turn there is a usage line: turn tokens (↑ input incl. cache / ↓ output) and the accumulated session cost in $ (on a subscription this is an estimate, not a separate bill).
• Screenshots the agent takes along the way are shown right in the chat feed (click to expand). They are not saved to history (localStorage is finite).
• You can paste images from the clipboard (Cmd+V in the input, up to 5 per message) — the model sees them; only a marker remains in history.
• "Ask before acting" mode (checkbox above the input): reading (tabs, text, screenshots, console, network) proceeds without questions, while mutating actions — clicks/typing/navigation/JS/forms/closing tabs/file uploads — wait for an Allow / Deny card. The agent sees a denial and continues the conversation. Toggling applies immediately, without recreating the session (via the Agent SDK's canUseTool).

On-page indication

When Claude acts on a tab (from the panel or from Claude Code):

• an orange glow burns around the page edges with a "✳ Claude is working…" badge, fading 2.5s after the last action;
• a virtual cursor (an orange arrow) glides to the action point and pulses a ring on click; it disappears after 3.5s of inactivity.

Both are hidden on screenshots so they don't end up in the frame and confuse the model when working with coordinates. On pages where scripts cannot be injected (chrome:// and the like) the indication is silently skipped.

Installation

1. Extension: clone this repository, open chrome://extensions (in the right space/profile!), enable "Developer mode", click "Load unpacked", and pick the extension/ folder.

2. MCP server — either way:

   • via npm: claude mcp add -s user chromium-bridge -- npx chromium-bridge
   • from the clone: cd server && npm install, then claude mcp add -s user chromium-bridge -- node "$(pwd)/index.mjs".

   It loads at session start — restart your Claude Code session after installing the extension.

   For other MCP clients, add this to your config:

   {
     "mcpServers": {
       "chromium-bridge": {
         "command": "npx",
         "args": ["chromium-bridge"]
       }
     }
   }
   

3. Check: the browser_status tool should return {"connected": true}.

Tools (v0.5)

Tool	What it does
browser_status	Check the connection to the extension
browser_tabs_list	List tabs (id, title, URL)
browser_tab_create / browser_tab_close	Open / close a tab
browser_navigate	Navigate to a URL; back/forward for history
browser_page_text	Page title, URL, and visible text
browser_computer	Mouse/keyboard/screenshots via CDP: clicks by coordinates or ref, drag, hover, type, key combos, scroll, zoomed region screenshot, wait
browser_read_page	Accessibility tree with ref ids (filter=interactive)
browser_find	Find elements by text/role, returns refs
browser_form_input	Set input/textarea/select/checkbox/contenteditable value by selector or ref
browser_click	DOM click by CSS selector (plain .click())
browser_upload_file	Put files into an <input type="file">
browser_javascript	Run JS on the page (await supported)
browser_console_messages	Tab console (with a regex filter)
browser_network_requests	Tab network requests (with a regex filter)
browser_resize_window	Window size
browser_gif_start / browser_gif_stop	Record a GIF of the tab → file; on long recordings the frame rate halves automatically, so the whole scenario fits

Everything except basic tab operations works through chrome.debugger (CDP): screenshots don't require activating the tab, clicks are real mouse events, and console/network are collected from the first CDP touch of the tab.

Limitations

• The extension lives in one browser profile — install it in the one you want to automate.
• While the server is running, its periodic ping keeps the extension's service worker awake. If the worker is asleep anyway (e.g. the server has just started), a keepalive alarm wakes it within ~30 seconds, and the server waits up to 12 seconds for reconnection before erroring.
• Trust model: the WS server listens on 127.0.0.1 and rejects connections whose Origin is not chrome-extension://…, which keeps web pages out. It does not distinguish between extensions, and a non-browser local process can fake the Origin header — anything running as your user is trusted, like with most local dev tools. Don't run the bridge on a shared machine.
• On the first CDP action the browser shows a "Chromium Bridge started debugging this browser" bar — that's normal, the debugger is the control mechanism. Closing the bar detaches the debugger (the next action re-attaches it).
• Console/network are not recorded retroactively — only after the tab is first touched.
• Port 8929 is owned by one session: a second parallel Claude Code session cannot start its own WS server (the extension stays with the first one).

License

MIT