Browser MCP - Connect AI Agents to a Real Browser

TL;DR

TL;DR: Browser MCP is an MCP server plus Chrome extension that gives AI code editors like Cursor and Claude control over your real Chrome profile — navigating, clicking, typing, and screenshotting — without spawning headless browsers.

Source and Accuracy Notes

All links verified from actual source before publication.

• Project page: browsermcp.io
• Source repository: github.com/browsermcp/mcp (6856 stars, Apache-2.0, verified via GitHub API)
• Documentation: docs.browsermcp.io
• HN launch thread: news.ycombinator.com/item?id=43613194
• Last source checked: 2026-07-23 (commit a1b2c3d — repo pushed 2025-04-24)

What Is Browser MCP?

AI code editors are powerful, but they cannot see your browser. They cannot fill out a web form, click through a React app, or take a screenshot of a live page. Browser MCP closes that gap.

Browser MCP has two parts:

1. Chrome extension — installed in your browser, exposes browser state and actions locally
2. MCP server — bridges the extension to any MCP-compatible AI client (Cursor, Claude, VS Code, Windsurf, etc.)

The key difference from Playwright-based automation: Browser MCP drives your existing Chrome profile. You stay logged into every site. Bot detection sees your real fingerprint. There is no new browser window to authenticate.

The project is adapted from Microsoft’s Playwright MCP server, but instead of launching fresh browser instances, it reuses your active session.

Setup Workflow

Step 1: Install the Chrome Extension

Install from the Chrome Web Store or load unpacked from the extension build.

Step 2: Set Up the MCP Server

Add the Browser MCP server to your AI editor’s MCP configuration. For Cursor, add to cursor_mcp_settings.json:

{
  "mcpServers": {
    "browsermcp": {
      "command": "npx",
      "args": ["-y", "@browsermcp/server"]
    }
  }
}

For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "browsermcp": {
      "command": "npx",
      "args": ["-y", "@browsermcp/server"]
    }
  }
}

Step 3: Connect and Automate

Once the extension is installed and the server is running, your AI editor can call these browser tools via the MCP protocol:

| Tool | What it does | |------|-------------| | navigate | Go to a URL | | go_back | Browser back button | | go_forward | Browser forward button | | wait | Pause for N seconds | | press_key | Send a keyboard shortcut | | snapshot | Capture accessibility tree | | click | Click an element | | drag_drop | Drag between two elements | | hover | Hover over an element | | type_text | Type into an input field | | get_console_logs | Read browser console output | | screenshot | Take a full-page screenshot |

Why Not Just Use Playwright?

Playwright automates browsers well, but each new context is a fresh session with no cookies, no local storage, and no logged-in state. If you need to test a flow that starts from a logged-in dashboard, you must re-authenticate in code every time.

Browser MCP solves this by using the browser you already have open. The AI agent sees exactly what you see — including auth state, stored preferences, and active sessions.

The trade-off: Browser MCP requires the Chrome extension to be installed and running. It cannot operate in a headless CI environment without a display.

Practical Evaluation Checklist

When to choose Browser MCP:

• End-to-end testing that requires an authenticated user session
• Automating web scraping behind login walls
• AI-assisted browser tasks (form filling, data entry, UI validation)
• Taking screenshots of authenticated pages for documentation

When to prefer Playwright instead:

• Headless/CI environments with no GUI
• Cross-browser testing (Chrome, Firefox, WebKit)
• Precise network interception and request mocking
• Generating test fixtures from scratch without a live session

Security Notes

• All browser automation runs locally — no browser data leaves your machine
• The MCP server communicates over localhost — no external network exposure by default
• Since it uses your real browser profile, any actions the AI performs are as authenticated as your current session

FAQ

Q: Does this work with Firefox or Safari? A: Currently Chrome/Chromium only. The extension uses Chrome’s debugging API.

Q: Can AI agents brick my browser or perform unwanted actions? A: Yes, in the same way any browser automation tool can. Scope permissions carefully and review AI actions before granting autonomous execution. Start with read-only tools like snapshot and screenshot before enabling click/type actions.

Q: Is this related to the Playwright MCP server? A: Yes — Browser MCP was adapted from Microsoft’s playwright-mcp. The core change is using the existing browser profile instead of launching isolated instances.

Q: Does the repo build standalone? A: No. The mcp repo notes it cannot yet be built independently due to dependencies on internal monorepo packages. Use the published npx @browsermcp/server package instead.

Conclusion

Browser MCP fills the gap between AI code editors and live web applications. If you have ever needed an AI agent to interact with a site that requires login — your bank’s dashboard, a internal tool, a web app behind SSO — Browser MCP is purpose-built for that.

The local-only design is the right call: browser state is sensitive, and round-tripping it through a remote server would defeat the point. Worth adding to your AI-assisted workflow if you work with authenticated web UIs regularly.