Browser MCP
Control your real Chrome from Claude Code.
29 tools. CAPTCHA solving. Multi-session. Human-in-the-loop.
$ npx @agent360/browser-mcp installCAPTCHA solving
Auto-detects reCAPTCHA v2/v3, hCaptcha, Turnstile, FunCaptcha. ~80% pass rate on checkbox CAPTCHAs via your real Chrome's trust score. Fallback to human-in-the-loop.
Real Chrome, not headless
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 10 concurrent Claude conversations against the same Chrome. Each gets its own Tab Group. No state corruption.
Human-in-the-loop
browser_ask_user shows an overlay dialog and blocks Claude until you respond. 2FA, credentials, CAPTCHA fallback: never in the transcript.
9 provider integrations
Stripe, HubSpot, Slack, Shopify, Pipedrive, Calendly, Mailchimp, Google, LinkedIn. Token extraction via browser_extract_token.
29 tools total
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 auto-snapshots, no accessibility trees per call. Playwright MCP floods your Claude context with page dumps after every click; Browser MCP doesn't. You can run 50+ actions before hitting a chat restart.
FAQ
How is this different from Playwright MCP?+
Playwright MCP runs a fresh headless browser per task: no logins, no cookies, no sessions. Browser MCP controls your real running Chrome via the Chrome Debugger API. Your Stripe, HubSpot, Calendly sessions stay intact. Also: Playwright MCP cannot solve CAPTCHAs or ask you for input mid-task. Browser MCP can.
How does CAPTCHA solving work?+
Auto-detects reCAPTCHA v2/v3, hCaptcha, Turnstile, FunCaptcha. For reCAPTCHA v2 the default action is click_checkbox, which passes ~80% of the time when you are signed into Google in that Chrome profile. Fallbacks: click_grid (AI vision guided) and ask_human (overlay dialog that pauses Claude so you can click three buses).
What does multi-session mean in practice?+
Each Claude Code conversation gets its own MCP server on a unique port (9876-9885). The Chrome extension creates a separate Tab Group per session with a distinct color. You can run 10 concurrent Claude agents against the same browser without tab conflicts or state corruption.
Is this safe? What can the extension actually see?+
The extension has Chrome Debugger API access, scoped to tabs you open or that Claude navigates to. No persistent background scraping. All commands flow through a local WebSocket (127.0.0.1): nothing leaves your machine. Source is open on GitHub. The browser_ask_user overlay is sandboxed so credentials you type do not appear in the conversation transcript.
Does it work with Cursor, Cline, and Claude Desktop?+
Yes. The npx installer writes the MCP config for Claude Code, Claude Desktop, Cursor, and Cline automatically. Other MCP-compatible clients work too, just point the client at npx -y @agent360/browser-mcp.
Is it free?+
Fully free and MIT-licensed. Both the npm package and the Chrome extension.
Does Browser MCP fill up my Claude context like Playwright MCP?+
No. Playwright MCP returns a full accessibility-tree snapshot after every navigate and click by default, which floods Claude's context and forces you to restart the chat on long tasks. Browser MCP returns only what you asked for: navigate gives you {title, url}, click gives you {ok}, screenshot gives you an image only when you explicitly call browser_screenshot. You can run 50+ actions in one conversation before hitting context limits.
Built by Agent360
We build voice AI and automation infrastructure for Nordic sales teams. Browser MCP is open-source because infrastructure should be shared.