It brings no data source of its own - the searching is the client's. What it adds is structure: forcing the question to be decomposed before searching, and the sources to be cited in the write-up. That is what separates a research report from a summary of the first five results.
A prompt-shaped server rather than a tool-shaped one. It ships a single deep-research prompt that walks a research question through a fixed sequence, using the client's own web search to gather sources.
- Question elaboration - the scope, the key terms and what is actually being asked
- Subquestions generated to cover the different aspects of the topic systematically
- A targeted search per subquestion, using the client's built-in web search
- Findings synthesised across sources, with citations kept attached
- A structured report produced as an artifact, with evidence-based conclusions rather than a link dump
Uv, and a setup script is provided for macOS. The name the README hands to uvx, mcp-server-deep-research, is not on PyPI. No API key of its own - the searching is done by the client, so it needs a client whose own web search is available. In use you select the deep-research prompt and give it a question.