Anyone can hand a model a search endpoint; what you get back is a naive keyword query. The systematic-review and PICO prompts push it to build the query a librarian would — MeSH terms, synonyms, date filters, a structured clinical question — which is the difference between a plausible reading list and a defensible one. Full text arrives as XML on purpose, so the model can tell a methods section from a discussion.
An MCP server over PubMed through NCBI's Entrez API, with three parts: search tools, abstract and metadata access via resource URIs, and prompts that help build a good query rather than a naive one.
- Search across the database using keywords, MeSH terms, author names, date ranges and Boolean operators
- Full text downloaded where it is available — for open access articles in PubMed Central — returned as XML rather than flattened prose, because the document structure is useful to a model
- Article abstracts and metadata retrieved through resource URIs
- Prompts that construct the query for you: a systematic review search with MeSH terms, synonyms and date filters; a PICO-framework clinical question; and a properly formatted author search
- Straight talk about the limits — the README says a failed full-text fetch does not mean the text is unavailable, and that paywalled papers are out of reach
Python, and `pip install mcp-simple-pubmed` (0.1.16), or a Smithery one-liner for Claude Desktop. `PUBMED_EMAIL` is required — NCBI asks for a contact address. `PUBMED_API_KEY` is optional and raises the rate limit from 3 requests per second to 10; the server implements no rate limiting of its own on the grounds that typical assistant use will not approach the cap. On macOS an SSL certificate verification failure is fixed by running the `Certificates.command` script that ships with a python.org install, then restarting the client. Licensed MIT.
One command plus a key — pip install mcp-simple-pubmed, then supply credentials