browser-use-mcp-server
Let AI agents control a real web browser to browse, navigate, and extract data via browser-use + Playwright.
Add to your client
Copy the config for your MCP client and paste it into its config file.
uv tool install browser-use-mcp-serverPaste into ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"browser-use-mcp-server": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"http://localhost:8000/sse"
]
}
}
}Claude Desktop connects to remote servers through the `mcp-remote` proxy (installed on first run via npx). Restart Claude Desktop after saving.
Step-by-step guides: Add to Claude Desktop · Add to Cursor · Add to Windsurf
Before you start
- uv (Python package manager)
- Playwright with a Chromium install (uv run playwright install --with-deps --no-shell chromium)
- mcp-proxy (required for stdio mode)
- An OpenAI API key (OPENAI_API_KEY)
About browser-use-mcp-server
browser-use-mcp-server exposes the browser-use automation framework over the Model Context Protocol, letting AI agents control a real Chromium browser through Playwright. Clients connect over SSE (uv run server --port 8000, then http://localhost:8000/sse) or over stdio using mcp-proxy. Browser tasks run asynchronously, and the session can be watched live via VNC. The agent is powered by an OpenAI model, so an OPENAI_API_KEY is required. A Docker image bundles the server together with a VNC server for an isolated, reproducible environment.
Tools & capabilities (1)
browser task (natural language)Accepts a natural-language instruction and drives a real Chromium browser to carry it out — for example, 'open https://news.ycombinator.com and return the top ranked article'. Browser operations execute asynchronously; set PATIENT=true if the API call should wait for the task to complete before returning. (The README documents the capability and example usage but does not enumerate individual tool names.)
What this server can do
browser-use-mcp-server provides tools for these capabilities — tap one to see every MCP server that does the same:
When to use it
- Browsing the web and extracting information directly from Cursor, Windsurf, or Claude Desktop
- Automating multi-step web navigation tasks driven by an AI agent
- Running browser automation in an isolated Docker container with live VNC viewing
- Scraping or summarizing page content (e.g., fetching the top-ranked article on a site)
Security notes
Requires an OPENAI_API_KEY supplied via a .env file or the client's env block. The Docker image exposes a VNC server (port 5900) with a default password of `browser-use`; override it by mounting a custom password file as a secret. The server drives a real browser that can visit arbitrary URLs, so treat tasks and target sites as untrusted.
browser-use-mcp-server FAQ
Which transports are supported?
Both SSE and stdio. SSE runs the server directly (`uv run server --port 8000`) and clients connect to `http://localhost:8000/sse`. stdio mode requires mcp-proxy and runs `browser-use-mcp-server run server --port 8000 --stdio --proxy-port 9000`.
What API key do I need?
An OPENAI_API_KEY, provided either in a .env file or via the client config's env block. Optionally set CHROME_PATH and PATIENT (true to wait for task completion).
Can I watch the browser while it works?
Yes. The server supports VNC streaming. The Docker image exposes VNC on port 5900 (default password `browser-use`), which you can view with a VNC client or noVNC.
How do I install it for stdio mode?
Build a wheel with `uv build`, then `uv tool install dist/browser_use_mcp_server-*.whl`. The published package is also available on PyPI as browser-use-mcp-server.
Alternatives to browser-use-mcp-server
Compare all alternatives →
All-in-one web access MCP — Web Unlocker, SERP, Scraper API, and a cloud Scraping Browser.
Popular community Playwright + API testing MCP server with codegen, screenshots, and device emulation.
Official Browserbase cloud-browser MCP built on Stagehand — natural-language act/extract/observe.
Compare browser-use-mcp-server with: