Skip to main content

Open source from Agent360

Browser MCP

You ask your AI assistant for a task on the web. Browser MCP lets it use your own Chrome, where you are already logged in, while you watch. If it needs a code, or there is a choice only you can make, it asks you first.

For example, it can:

  • Fetch this month’s figures from a system you are already logged in to, and put them in a table
  • Fill in a form and ask you before it does anything sensitive
  • Log in, fetch the one-time code from your email and carry on

The extension is installed in Chrome on a computer.

users in the Chrome Web Store
2,000
users in the Chrome Web Store
downloads from npm since April
20,090
downloads from npm since April

Measured 26 October 2026.

The next tool is on the way

We’re releasing more of our own tools as open source. There’s more to come from us.

Follow along on GitHub →

Do you want to use AI agents in your own systems? Write to us →

Animated film, 20 seconds. The tab groups are as in the real extension. The pages and the log are drawn.

For developers

Open source · MIT licensed

Browser MCP

Control your real Chrome from Claude Code.
CAPTCHA help. Multi-session. Human-in-the-loop.

$ npx @agent360/browser-mcp install

CAPTCHAs

Auto-clicks the reCAPTCHA v2 checkbox, with a human fallback for the rest.

Your real Chrome

Uses the Chrome Debugger API against your running browser. Your logins, cookies, 2FA sessions: all intact. No re-authentication per task.

Multi-session isolation

Up to 20 Claude conversations against the same Chrome. Each gets its own colour-coded tab group.

Human-in-the-loop

browser_ask_user shows an overlay dialog and waits for your answer (two minutes by default): for 2FA, logins and CAPTCHAs.

9 provider shortcuts

browser_extract_token opens a provider page and returns instructions for finding or creating API credentials. It includes shortcuts for Stripe, HubSpot, Slack, Shopify, Pipedrive, Calendly, Mailchimp, Google and LinkedIn.

Everyday browser tools

Navigate, click, fill, screenshot, file upload, console logs, cookies, localStorage, iframe support, network waits, page content: everything you need.

Minimal context footprint

Every tool returns only what you asked for: no automatic page snapshots, and an image only when you call browser_screenshot.

FAQ

How is this different from Playwright MCP?+

Both can work with a logged-in browser: Playwright MCP keeps a persistent profile by default and can connect to your Chrome through its extension. The difference we have measured is human-in-the-loop: Playwright MCP has no tool that stops and asks the person, and Browser MCP has browser_ask_user for 2FA codes, CAPTCHAs and choices only you can make.

How does it handle CAPTCHAs?+

The default action is detect: it finds the CAPTCHA on the page. For reCAPTCHA v2, click_checkbox tries the checkbox and click_grid clicks the cells you choose (it returns no image, so take a screenshot first). For the rest, ask_human returns a message to show with browser_ask_user, so a person finishes it in your own Chrome.

What does multi-session mean in practice?+

Each Claude Code conversation gets its own MCP server on a free port (9876-9895). The Chrome extension gives each session its own colour-coded tab group in the same Chrome profile. Up to 20 sessions can run at once.

Is this safe? What can the extension actually see?+

The extension uses the Chrome Debugger API, and the agent works only in the tabs its own session opened. No persistent background scraping. The MCP server and the extension talk over a local WebSocket (127.0.0.1). Page content Claude reads, and anything you type into a browser_ask_user dialog, is sent to your AI client. Source is open on GitHub.

Does it work with Cursor, Cline, and Claude Desktop?+

Yes. The npx installer registers the server with Claude Code, Codex, VS Code and Cursor when it finds them. Claude Desktop, Cline and other MCP clients work too: add npx -y @agent360/browser-mcp@latest to that client's MCP config.

Is it free?+

Fully free and MIT-licensed. Both the npm package and the Chrome extension.

How much of my Claude context does it use?+

Browser MCP returns only what you ask for: navigate returns the page title and URL, click returns whether it landed, and you get an image only when you call browser_screenshot. How much context a task uses depends on the tools, the page content and your AI client.

Built by Agent360

We build AI infrastructure for Danish companies. Browser MCP is open-source because infrastructure should be shared.