Returning the AI answer and the source list together is the useful shape: the model can lead with the summary and still quote the page it came from. It is a thin wrapper on one vendor, so your search bill and rate limits are Tavily's.
A small MCP server that hands a query to Tavily and returns the results as text. Each response carries an AI-generated answer alongside the source titles and URIs, so the model gets both a summary and something to cite.
- `search` — run a query against Tavily and get back the results as text, with the AI response, URIs and titles of the matches
- A `search_depth` argument set to either basic or advanced, so a quick lookup and a deeper pass are the same tool with a different setting
A Tavily API key, supplied as `TAVILY_API_KEY`. Run from a clone: the documented client entry uses `uv --directory <path> run tavily-search`, with `PYTHONIOENCODING` set to utf-8 to keep non-ASCII results intact. A Docker Compose file is included for hosts where the desktop clients are unavailable.