Summary

One tool, two useful outputs: an answer and the sources it came from.

Getting the synthesised answer and the underlying URLs in the same response is what makes this usable for real questions — the model can answer straight away, and you can still click through when it matters. The depth setting is the only knob: basic for a quick lookup, advanced when the first pass was too shallow.

What it is

A small Python MCP server that runs searches through the Tavily API and returns the results as text — an AI-generated answer first, then the titles and URLs behind it.

What you get

One tool, `search`, taking a query and an optional search depth of basic or advanced
An AI answer synthesised from the results, so the assistant has something to work with immediately
The source titles and URLs alongside it, so claims can be checked rather than taken on trust
Server-side logging of each search and its outcome, written to a per-platform log directory

Requirements

A Tavily API key set as `TAVILY_API_KEY`, plus uv to run the cloned repository. On Windows also set `PYTHONIOENCODING` to utf-8. For Cursor, wrap the launch in a shell script that exports the key and starts the server. A Docker Compose setup is provided for Linux, where the desktop client is not available.

Setup effort

One command plus a key — npx -y @smithery/cli install tavily-search --client claude, then supply credentials