Purpose real-browser-mcp is aimed at the coding and verification loop: when you are already logged in in a real Chrome, already on a staging environment, and already on the page where a bug has been reproduced, let the AI agent directly operate this browser to verify the result, instead of opening a separate brand-new browser with no cookies. The project is positioned as a bridge solution of "MCP server + Chrome extension": the AI client sends instructions via the MCP protocol, the real Chrome is responsible for execution, and session data stays local.
Main capabilities It provides 18 tools, grouped by purpose, supporting operations such as clicking, typing, navigation, screenshots, and snapshots. Connection interruptions are automatically retried (exponential backoff from 1 second to 30 seconds), a ping/pong health check is performed every 10 seconds, and timeouts are set for different tools (5 seconds for clicking, 60 seconds for navigation). It supports running two server instances simultaneously (on different ports). It comes with a configuration file that can teach the agent the workflow of "snapshot first, then act, then verify," and handles the choice between dropdown menus, screenshots, and snapshots. The repository is published as an Agent Plugins 1.0.0 package, containing plugin.json, mcp.json, and skills/real-browser-control/.
How to use
- Add the MCP server: you can install it with one click in Cursor, or manually in Cursor Settings > MCP > Add new MCP server; for Claude Desktop, edit the configuration file claude_desktop_config.json (macOS path ~/Library/Application Support/Claude/, Windows path %APPDATA%\Claude); for Windsurf, add the same JSON configuration in Settings > MCP. Any MCP-compatible client works.
- Install the Chrome extension: when loading from source, open chrome://extensions, enable Developer mode, click "Load unpacked," select the extension/ folder in the repository, and click the Real Browser MCP icon in the toolbar. A green dot means connected, gray means waiting for the server.
- Cursor local plugin: after cloning the repository, copy or symlink it to ~/.cursor/plugins/local/real-browser-mcp and reload the window; you still need to connect the Chrome extension.
- Teach the agent the workflow: run a command to generate the rules file, then in the conversation enter /check-browser or say "check the result in my browser"; Claude Code will automatically discover AGENTS.md in the project root.
Applicable scenarios Suitable for debugging and verification scenarios in a staging environment where you are already logged in, on pages where enterprise SSO has already been completed manually, and where real-time session state is needed. The project explicitly recommends: use Playwright when repeatable CI automation is needed; use this tool when authenticated real-time state in the browser is needed.
Limitations The Chrome extension must be installed; the MCP server alone cannot see or control Chrome. The extension runs in a normal Chrome profile, so the agent can click, type, and read visible content in connected tabs (including logged-in applications). When testing untrusted websites, it is recommended to use a dedicated profile or tab, and do not point automation at tabs containing sensitive data unless you accept that risk. This tool is not an "agent browser" (it does not create a new headless/cloud browser), nor is it a replacement for Playwright or Chrome DevTools MCP; Chrome 136+ has blocked remote debugging of the default user profile, so this tool instead uses a localhost WebSocket extension bridge. All communication takes place locally, with no cloud and no proxy, and the page content returned by the tools will still enter your AI client. Details such as price, fees, and platform support scope are not provided in the materials.
