The trade is deliberate: instead of the model choosing click, type and scroll one at a time, it hands over one instruction and the browser-use agent works out the steps. That keeps the tool surface tiny and the token cost low, and it means you get less control over how the task is carried out.
An MCP wrapper around the browser-use automation engine. It exposes a single tool that opens a browser session, runs the agent against your instruction, and returns the final result - the browser lifecycle and cleanup are handled inside.
- One tool, run_browser_agent, that takes a natural-language task and reports back what happened
- Navigation, form filling, tab control, screenshots and page reading, driven by the browser-use agent
- Browser configuration from environment variables - persistent profiles, proxies and custom Chromium flags - without touching agent code
- A choice of model provider and model, with step and action limits capped per run
- Redaction of proxy settings and debug endpoints from logs, since those can carry credentials
Python 3.11+, Google Chrome or Chromium for local automation, and uv for dependency management. Clone, run uv sync, then launch with uv run. Configuration is environment variables: the model provider and name, the provider's own API key, and the browser runtime flags. The project's own security note is the one to read first - driving a full browser grants broad access to the host machine.
Build from source — clone the repository and build it, then point your client at the binary