The browsing is driven by the provider you configure, not by the assistant that called the tool — so one request can turn into a long chain of model calls on your key. That is also the appeal: a multi-step task completes without your assistant narrating every click. Pointing it at Ollama keeps the whole loop local if the cost matters more than the capability.
A FastMCP server that wraps browser-use, so an assistant can delegate a whole browsing task rather than issuing clicks. You give it an instruction — navigate here, search for this, summarize what you find — and it runs the steps itself in a real browser, driven by a model you choose at launch.
- Browser automation from natural language: navigating, filling forms, clicking buttons and other web tasks carried out as a sequence, not one call at a time
- A real Chromium via Playwright, so pages that need JavaScript behave as they would for a person
- Your choice of model behind it, set with `--model` at launch
- A wide provider list — OpenAI, Anthropic, Google, Cohere, Mistral, Groq, Together, AWS Bedrock, Fireworks, Azure OpenAI, Vertex AI, NVIDIA, AI21, Databricks, IBM watsonx.ai, xAI, Upstage, Hugging Face, and local options through Ollama or Llama.cpp
- `--debug` for detailed logging when a run does something you did not expect
An API key for whichever provider you pick — `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `GROQ_API_KEY` and so on, or `OLLAMA_BASE_URL` to stay local. Install the package with the extra for your provider, or `[all-providers]` for everything, then `playwright install chromium`, which is the step most often skipped. The client config runs `browser-use-mcp` with `--model` in `args` and the key in `env`; set `DISPLAY` on a GUI environment. Package `browser-use-mcp` at 0.1.0, MIT.
One command plus a key — pip install -e "git+https://github.com/pietrozullo/browser-use-mcp.git#egg=browser-use-mcp[openai]", then supply credentials