The absence of an API key is the whole point: search and extraction both run locally, so there is no vendor account, no quota and no per-query bill to think about. The trade-off is the usual one for scraping — sites that block automated fetches will come back empty, and rephrasing the query is the practical fix.
A research tool that runs entirely on your machine. It takes a question, searches DuckDuckGo, fetches the top results, strips each page down to readable text with Trafilatura, and hands the extracted content back to the model. Nothing is sent anywhere except the public web searches themselves.
One research call taking a `query` string and a `num_results` count — three by default — that searches, scrapes and extracts in one step, so the model receives article text rather than a list of links it cannot open. Requests are spaced out politely and queries are sanitised before they are used.
Nothing — no account, no key. Python 3.10 or higher and `uv`; from a clone, `uv add "mcp[cli]" duckduckgo-search trafilatura` installs the dependencies. The documented client entry runs `uv --directory <path> run python research_server.py` over stdio.