Browser Use MCP Server

What co-browser/browser-use-mcp-server MCP server does

co-browser/browser-use-mcp-server MCP server exposes browser-use browser automation through the Model Context Protocol. An MCP client can send browser tasks to an instance running locally or in a container. The README demonstrates a task that opens Hacker News and returns the top-ranked article, indicating that the server is intended for navigation and web interaction driven by an AI agent.

The project supports two MCP transports. SSE mode exposes the server at an HTTP endpoint, while stdio mode runs through the installed command and uses mcp-proxy as part of the stdio setup. Browser work can also run asynchronously, allowing operations to continue without requiring every request to block until completion.

How it works

The server uses browser-use together with Playwright to automate Chromium. In local development, dependencies are synchronized with uv, Playwright is installed, and Chromium is downloaded with its required dependencies. The server can then be started from source with an HTTP port, or built into a wheel and installed as a uv tool for stdio operation.

SSE clients connect to the /sse path on the configured server port. The stdio example starts browser-use-mcp-server with the run server arguments, enables stdio, and assigns a proxy port. Client configuration examples are provided for Cursor, Windsurf, and Claude Desktop.

For containerized operation, the included Dockerfile packages the browser environment and VNC support. Port 8000 is used for the MCP server and port 5900 for VNC. A browser-based noVNC viewer can connect to the VNC port so an operator can watch the automation. The default VNC password is browser-use, unless a mounted password file overrides it.

Setup and configuration

The documented local prerequisites are uv, Playwright, and mcp-proxy. After installing dependencies, install Chromium with:

uv run playwright install --with-deps --no-shell chromium

Set OPENAI_API_KEY in a .env file or in the process environment before running the server. CHROME_PATH can point to an alternate Chrome executable. The PATIENT setting controls whether API calls wait for task completion; the README uses false by default.

SSE mode can be started from the source tree with uv run server --port 8000. For stdio, build the project, install the generated wheel as a uv tool, and run the installed browser-use-mcp-server command with the documented server and proxy arguments. Docker users can build the image locally and publish ports 8000 and 5900.

Tools and capabilities

co-browser/browser-use-mcp-server MCP server provides the following documented capabilities:

• Control web browsers through AI agents.
• Communicate through SSE or stdio MCP transports.
• Run browser operations asynchronously.
• Stream the browser session through VNC.
• Use Chromium locally or in the supplied Docker environment.

The provided material does not enumerate individual MCP tool names or their parameter schemas, so integrations should not assume a specific tool catalog beyond these capabilities.

Limitations and notes

An OpenAI API key is required by the documented environment configuration. Browser automation also depends on a working Playwright and Chromium installation unless the Docker image is used. Stdio mode requires mcp-proxy according to the prerequisites, while SSE mode is shown as a direct server process.

The Docker example uses a default VNC password unless one is supplied through the mounted secret file. Change that default before exposing the VNC port beyond a trusted environment. The README documents configuration paths for Cursor, Windsurf, and Claude Desktop, but does not provide a configuration example for every MCP client.