WEBBRIDGE agent-webbridge · MIT · runs on your computer
Open source · MIT · no account · no telemetry

WebBridge — let your AI use your own signed-in Chrome

Agent WebBridge (agent-webbridge) is a local browser-automation MCP server for Claude Code, Cursor and other AI apps. It drives the Chrome you already use, so your logins are already there, across several Chrome profiles and several tabs at once. It runs on your computer (127.0.0.1).

Easiest way: paste this into your AI

Open an AI assistant that can run terminal commands (for example Claude Code or Cursor), paste this, and follow along. You only click “Add to Chrome” yourself.

Please set up agent-webbridge for me so you can use my own Chrome (open source: https://github.com/jeet-dhandha/agent-webbridge). Do these steps one at a time and tell me what you did after each one:

1. Run `node -v`. If it is older than version 18, stop and tell me.
2. Run `npm i -g agent-webbridge`.
3. Run `awb profiles` and ask me which Chrome profile to use.
4. Run `awb setup "<that profile>"`. It opens the Chrome Web Store, so tell me to click "Add to Chrome", then wait for me.
5. Connect yourself with `claude mcp add chrome -- npx -y agent-webbridge mcp`. If you are not Claude Code, show me the MCP config to add instead: {"mcpServers":{"chrome":{"command":"npx","args":["-y","agent-webbridge","mcp"]}}}
6. Run `awb doctor` and tell me what it says.

Ask me before you close or restart Chrome. Do not run anything else. Everything stays on my computer (127.0.0.1).
What this does
  • Checks you have Node.js 18 or newer
  • Installs the agent-webbridge package
  • Opens the Chrome Web Store so you add the extension
  • Connects your AI to it, then runs a self-check
Read before you paste. The prompt runs six commands and asks before closing Chrome. With an up-to-date extension, connecting does not restart Chrome; an older build may need Chrome closed, so save your work first. Not for a plain chat window that cannot run commands.
Prefer to do it yourself?
npm i -g agent-webbridge awb setup "Work" claude mcp add chrome -- npx -y agent-webbridge mcp

Cursor, Claude Desktop or Windsurf: add {"mcpServers":{"chrome":{"command":"npx","args":["-y","agent-webbridge","mcp"]}}} to the MCP settings. Needs Google Chrome and Node.js 18+, on macOS or Windows.

What can I ask it to do?

Shown in demo 1

Read a page and summarize it

Open https://news.ycombinator.com, read the top 10 stories, and give me a 5-line summary.
Shown in demo 2

Compare several pages at once

Open these 5 pages in separate tabs at the same time and tell me the main difference between them: <url 1> <url 2> <url 3> <url 4> <url 5>
Shown in demo 3

Work inside a site you are logged into

In my Chrome profile "Work", open <the site>, fill in the new-item form with <my details>, and stop before you submit. Show me what you filled in.
Shown in demo 4

Fill in a form

Fill the form at <url> with my name, email and message, choose the right options, and do not press submit.
Recipe

Upload a file

On <url>, attach the file ~/Documents/<file>.pdf to the upload field and tell me when it is attached. Do not submit.
Multi-profile guide

Use more than one Chrome profile

Open <the dashboard> in my "Work" profile and my "Personal" profile and tell me which one has unread alerts.
You stay in charge. The AI acts as you in your own browser. Ask it to stop before anything is submitted or published, and review what it did.

Execution Demonstration (Real Runs)

Recorded from live MCP runs driving authentic Chrome

Real runs, not mock-ups: an AI agent sends MCP calls to npx agent-webbridge mcp, which drives a real, signed-in Chrome profile. Timings are measured live on a MacBook on broadband. Public pages only; nothing is submitted to any account.

1. Read Any Page in Your Own Chrome
Snapshot + Eval
Real demo: agent-webbridge taking accessibility snapshot and evaluating JavaScript on Hacker News
Snapshot + evaluate on Hacker News: extracts structured accessibility trees with stable @e element references, read from your own signed-in Chrome.
2. True Parallelism (5 Tabs Concurrently)
3.5s vs 9.3s (2.6x Speedup)
Real demo: 5 parallel tabs evaluated concurrently in 3.5 seconds vs 9.3 seconds serial
Five tabs evaluate at once: the extension attaches chrome.debugger per tab (a Map keyed by tabId), finishing in 3.5s in parallel versus 9.3s serially.
3. Drive Real Authenticated Sessions
Zero Re-Auth · Preserves 2FA
Real demo: driving a signed-in editor safely using authentic cookies
Your real login sessions: a signed-in web app safely filled and verified with authentic cookies and localStorage state, stopping safely before publishing.
4. Fill & Submit Complex Forms
Inputs, Radios & Contenteditable
Real demo: filling native inputs, radio buttons, checkboxes, and contenteditable editors
Comprehensive input dispatch: fills native text inputs, radio selections, checkboxes, and modern contenteditable rich text editors via CDP input events.

Honest limits

  • It automates a browser you are already signed in to. Sites can still detect or limit automation, and you must follow their terms.
  • Timings are from one MacBook on home broadband; yours will differ.
  • macOS and Windows with Google Chrome and Node.js 18+. Chrome launching on real Windows is not verified yet (the test suite passes on Windows CI).
  • No record-and-replay yet: you describe the task, and community recipes hold tested steps for specific sites.

Learn more

Quick answers

What is WebBridge (agent-webbridge)?

WebBridge, published as agent-webbridge, is an open-source (MIT) browser-automation MCP server and Chrome extension. It lets AI agents such as Claude Code or Cursor use your own signed-in Chrome profiles, with several tabs and profiles running at the same time. Everything runs on your computer (127.0.0.1).

Do I need to know how to code?

Not to install it. Paste the setup prompt into an AI assistant that can run terminal commands, such as Claude Code or Cursor. You need Google Chrome and Node.js 18 or newer, and you click "Add to Chrome" yourself. A plain chat window that cannot run commands or connect MCP servers cannot set it up.

Is it safe to let an AI use my logged-in Chrome?

It is powerful, so treat it like handing someone your browser. Since version 1.3.1 the bridge refuses requests that come from web pages or from non-local hosts, and versions before 1.3.1 had a security hole, so update. The AI acts as you: watch what it does, ask it to stop before submitting anything important, and use it only on accounts and sites you are allowed to automate.

Which AI apps work with it?

Any app that supports MCP servers, or that can send HTTP requests to the local router. Setup instructions are written for Claude Code, Claude Desktop, Cursor and Windsurf.

Can it use several Chrome profiles at once?

Yes. Each profile gets its own local daemon, and a router sends each request to the profile you name. Inside a profile every tab is controlled separately, so tabs run at the same time. In our benchmark on one MacBook, 5 tabs finished in 3.5 seconds in parallel versus 9.3 seconds one at a time; results vary by machine and site.

How is it different from Playwright or a cloud browser?

Playwright and cloud browsers usually start a fresh browser, so you log in again and deal with 2FA. WebBridge uses the Chrome you already use, so existing sessions are reused. It does not make automation undetectable: sites may still limit automated behavior, and you must follow their terms.

Does it record my clicks and replay them?

No, there is no recorder yet. You describe the task to your AI, and community recipes (awb-recipes) capture tested steps for specific sites.