- 1 · Nothing launches Your existing signed-in tab answers. No fresh window animates in — the absence is the proof.
- 2 · It fills a real form
fill_formon the live, logged-in page you were already looking at. - 3 · Two at once Two terminals driving two Chrome profiles simultaneously, each addressed separately.
Why other Chrome MCP servers open a fresh profile you're not signed into
It isn't laziness on their part. It's a lock Chrome itself put in place.
Most browser MCP servers drive Chrome over the DevTools Protocol (CDP), which needs Chrome started with a remote debugging port. Since Chrome 136, a Chrome running your real user profile will not also hand that port to an external CDP client — doing so would compromise the Safe Storage HMAC that protects your cookies and saved passwords. Chrome removed the combination on purpose.
"It's no longer possible to use the default profile, chrome removed the feature intentionally in >=v136."
— browser-use's maintainer, on why it can't use your real Chrome profile. Checked 2026-07-17.
That's why every CDP-based MCP server — playwright-mcp, chrome-devtools-mcp, browser-use — either copies your profile, launches a fresh one, or asks you to quit Chrome first. It isn't an oversight they'll patch. Chrome closed the door.
AgentHub isn't a CDP client. It's a browser MCP Chrome extension talking to a local bridge over Chrome's native messaging, which is why it can act inside the tab you already have open, in the profile you're already signed in to.
To be precise about what this does and doesn't buy you
This is a structural advantage over CDP-based tools, not over every tool. Other extension-based MCP servers — mcp-chrome and BrowserMCP — also reach your real profile, same architecture, same door. Against those two specifically, our difference isn't the profile: it's that several profiles can run at once, each addressed separately, which mcp-chrome's own #347 tracks as an open request. Running several clients at once isn't unique to us.
Run two Claude Code sessions, or work and personal Chrome profiles, at the same time
One bridge, many browsers. Each connection gets its own slot instead of overwriting the last one.
Several sessions. Several profiles. No last-connect-wins.
Each connected client gets its own slot, and every request is tagged with a browser-bound id — so a second Claude Code session never kills the first. Separately, each browser and profile gets its own namespaced ID (chrome:abc-123), so an AI driving your work and personal profiles never confuses them.
Your real browser, your real data
No fresh Chromium, no copied profile, no "please quit Chrome". Gmail is logged in. Slack is there. Your paid SaaS subscriptions work, because they're your actual sessions.
One bridge, every tool
AgentHub speaks standard MCP, so any MCP-compatible AI tool works without extension changes. Add a new AI to your stack tomorrow without reinstalling anything in the browser.
Privacy-first defaults
Banking and email domains are blocked by default. A real-time activity log shows every tool call. Every tool has its own on/off toggle.
AgentHub vs Browser MCP, mcp-chrome and chrome-devtools-mcp
Their own public words, quoted and linked to each project. Read them and decide for yourself.
chrome-devtools-mcp (Google)
"The current MCP server binds to a single Chrome instance for its entire lifetime."
— chrome-devtools-mcp issue #926. It's CDP-based and launches a fresh, dedicated profile by default. Checked 2026-07-17.
mcp-chrome
"second Claude Code session kills first via shared MCP Server singleton"
— mcp-chrome issue #345, open as of 2026-07-17. Multi-profile support is a separate open request, #347: "a common workflow for separating personal / work / per-project browsing". It's extension-based, so it does reach your real profile.
Browser MCP
"Close any existing connections"
— a string in Browser MCP's own source code: a new connection closes the previous one. Last push April 2025. Checked 2026-07-17.
playwright-mcp (Microsoft)
"A persistent profile can only be used by one browser instance at a time, so concurrent MCP clients sharing the same workspace will conflict."
— playwright-mcp's README. Not your real Chrome by default. Checked 2026-07-17.
browser-tools-mcp
"THIS PROJECT IS NO LONGER ACTIVE PLEASE USE A DIFFERENT SOLUTION FOR THIS."
— line 1 of browser-tools-mcp's own README. Checked 2026-07-17.
Every quote above is the project's own public text, linked to source. We're not claiming any of them is broken — several are excellent, and the extension-based ones reach your real profile exactly like we do. Read them and decide.
What Claude Code and Cursor can do in your real Chrome
Every example is a real prompt your AI can answer once AgentHub is installed.
Research, faster
Pull insight from any tab — articles, docs sites, search results, social profiles.
"Summarize this 60-page docs site." "Pull every job title from this LinkedIn profile into a table." "What's the pricing across all three competitor tabs I have open?"Forms, without copy-paste
Fill Salesforce, job applications, CRMs, contact forms — using data from another tab or a file.
"Fill this Salesforce lead with Jane's info from the tab next to me." "Apply to this job using my resume in tab 2." "Submit this contact form with the values from my CSV."Cross-tab orchestration
Your AI sees every tab, switches between them, and combines results into one answer.
"Find Jane's email in tab 1, her company's last funding round in tab 2, and fill them into the notes in tab 3."Drive your browser from your IDE
Claude Code or Cursor running in your terminal can ask the browser questions — and act on the answers — without leaving the session.
"Open this URL in Chrome and fill the form with the test data from my repo." "What's on the page I'm looking at right now?"Works with Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, JetBrains and Zed
One install, one bridge, standard MCP. Add a new AI to your stack tomorrow without touching the browser.
MCP bridge
The main event. No API key needed — these tools handle their own auth.
- Claude Code
- Claude Desktop
- Cursor
- VS Code
- Windsurf
- JetBrains IDEs
- Zed · Continue.dev
Optional chat sidebar
If you want a chat in the browser too. Bring your own API key, stored locally.
- OpenAI
- Anthropic Claude
- Google Gemini
- Custom model IDs
Browser MCP tools: get_page_content, take_screenshot, fill_form, extract_table
Every tool toggles on or off individually. Banking and email domains are blocked by default.
| Tool | What it does |
|---|---|
| get_page_content | Read text from any page |
| take_screenshot | Capture what you see |
| list_tabs | See every tab open across windows and profiles |
| get_page_metadata | Page title, description, OG tags |
| navigate | Open or move to a URL |
| fill_form | Auto-fill form fields (React, Vue, Lit, plain HTML, shadow DOM) |
| click_element | Click buttons, links, controls — by text, selector, or index |
| press_key | Submit forms or trigger keyboard shortcuts |
| extract_table | Pull HTML tables as structured rows |
| read_form | Inventory every input on a form |
| extract_data | Heuristic detection of structured data on a page |
What actually leaves your machine
You're about to give a tool your real cookies and every tab. Here is the honest inventory — including the one thing that is on by default.
| What | Where it goes |
|---|---|
| Page content, DOM, screenshots | Never sent Stays between the extension and the bridge on your machine, over Chrome native messaging. |
| Chat sidebar API keys | Never sent Stored in chrome.storage.local. They go to OpenAI, Anthropic or Google only when you send a message — never to us. |
| Redacted diagnostic logs | On by default Packaged builds upload redacted logs, keyed by a random per-install ID, to our ingest endpoint. Turn it off with remote.enabled: false in logs-config.json. |
| Open ports | One: TCP 7483, bound to 127.0.0.1. Loopback only, never exposed to your network. Verify with lsof -i :7483. |
This page is a normal web page — it runs Google Analytics and carries an ad script, like the rest of the site. The claims above are about the AgentHub product, not about this page. Details in the privacy policy.
Two steps, then a restart
Add the extension, run one npx command, restart your AI tool. No account, no waitlist, no API key needed for the MCP bridge.
Needs Node 18+ — check with node -v. Chrome 120+ on macOS, Windows or Linux.
Add the extension
Install AgentHub from the Chrome Web Store. Do this first — the setup command in step 2 finds the extension by looking for it in your browser.
Add to Chrome — FreeRun the setup command
It installs the bridge, auto-detects your extension ID, and registers AgentHub with every AI tool it finds — Claude Code, Cursor, VS Code, Windsurf, JetBrains, Zed, Continue. If it finds more than one AgentHub extension, or none, it stops and prints the exact command to re-run with --extension-id <id>.
npx agenthub-setup@latest --yes
Then restart your AI tool
Claude Code, Cursor and VS Code read their MCP config at startup — so a correct install looks like nothing happened until you restart them. Then ask your AI about the tab you have open.
Something not working? The support page has the troubleshooting steps.
FAQ
Can I use my existing logged-in Chrome profile instead of a new one?
Yes — that is the entire point of AgentHub. It's a Chrome extension plus a local bridge, so it drives the Chrome you already have open, with your cookies, your logins and your sessions. Nothing new launches, and you don't have to quit Chrome first.
Why does my MCP server open a new Chrome profile that isn't signed in?
Because it drives Chrome over the DevTools Protocol (CDP). Since Chrome 136, a Chrome running your real user profile won't also expose a remote debugging port to an external CDP client — that would compromise the Safe Storage HMAC protecting your cookies and saved passwords. So CDP-based tools have to copy your profile, launch a fresh one, or ask you to quit Chrome.
An extension doesn't use CDP, so it isn't subject to that restriction. That's the whole architectural difference — see the section above.
Can I run two Claude Code sessions against the same browser at once?
Yes. The bridge keeps a separate slot per connected client and tags each request with a browser-bound id, so two clients issuing the same JSON-RPC id can't clobber each other. A second Claude Code session doesn't kill the first.
Worth being straight with you: several other browser MCP servers also support multiple clients. This isn't unique to AgentHub — the multi-profile part below is the unusual bit.
Can Claude Code use my browser?
Yes. AgentHub is a browser MCP for Claude Code — add the Chrome extension, run the setup command, and Claude Code can read, click and fill the Chrome tab you already have open, signed in and all.
The same Chrome browser MCP bridge also works with Cursor, VS Code, Windsurf, JetBrains and Zed over the standard Model Context Protocol.
Can I connect my work and personal Chrome profiles at the same time?
Yes. Every connected browser and profile gets its own namespaced composite ID — for example chrome:abc-123 — and tab IDs returned by list_tabs are scoped to it, so an AI driving two profiles never confuses them.
Chrome, Edge, Brave, Arc and Vivaldi can be connected at the same time and addressed separately.
Does AgentHub open a port on my machine?
Yes — one. The bridge listens on TCP 7483, bound to 127.0.0.1 (loopback), so it's reachable only from your own machine and is never exposed to your network.
Don't take our word for it: lsof -i :7483.
What does AgentHub send to its servers, and how do I turn it off?
Page content, DOM, screenshots and your chat API keys are never sent anywhere — the extension talks to the bridge on your own machine over Chrome native messaging.
One exception, stated plainly: packaged builds do send redacted diagnostic logs by default, keyed by a random per-install ID. That's opt-out, not opt-in. Set remote.enabled to false in logs-config.json and nothing is uploaded. Full breakdown in what actually leaves your machine.
Is AgentHub open source?
The source is on GitHub at github.com/irmasemma/AIBrowserCopilot — you can read it, audit the manifest and build it yourself.
Being precise, because this audience cares: the repo doesn't carry a LICENSE file yet, so it's source-available, not open source under the OSI definition. A license is on the list, and we'll say "open source" when that's true and not before.
How is this different from Claude for Chrome?
Claude for Chrome is Anthropic's own in-browser assistant — Claude works inside your browser. AgentHub is the other direction: it gives the AI tools already running in your terminal or IDE — Claude Code, Cursor, VS Code, Windsurf, JetBrains, Zed — a browser to drive, over the standard Model Context Protocol.
If you want Claude in a sidebar, use Claude for Chrome. If you want the agent in your terminal to see and act on the tab in front of you, that's this.
Which AI tools work with AgentHub?
Out of the box: Claude Code, Claude Desktop, Cursor, VS Code (Copilot or Continue), Windsurf, JetBrains IDEs, Zed and Continue.dev. Any MCP-compatible AI tool works via the standard protocol — the bridge doesn't care which client is talking to it.
The extension also ships an optional chat sidebar that takes your own OpenAI, Anthropic or Gemini key.
Do I need an API key?
Not for the MCP bridge — Claude Code, Cursor and the rest handle their own auth. AgentHub just exposes the browser as a tool source for them.
An API key is only needed for the optional in-extension chat sidebar, where you bring your own OpenAI, Anthropic or Gemini key. It lives in chrome.storage.local and never leaves your machine.
Where do I report bugs or request features?
GitHub issues: github.com/irmasemma/AIBrowserCopilot/issues
Point your AI at the tab you already have open.
Free Chrome extension plus a one-command local bridge. Several sessions and several Chrome profiles at once — each addressed separately, so the second one never kills the first.
Add to Chrome — FreeNo account · Free · Bridge binds to 127.0.0.1 only