Cool project, full text dreaming...

[froggie901]

I think the idea of this kind of MCP is really cool. I'm new to the agentic space, but I think to unlock better responses overall. How do we get access to the actual information in the article, no the metadata and abstracts.

[cyanheads]

@froggie901, pubmed_fetch_fulltext is the tool you want, and as of v2.6.0 (shipped today) it covers a lot more than it used to.

If the article is in PubMed Central, you get the structured JATS body: sections, subsections, references. That path has been there since v1. New in v2.6.0: if the article isn't in PMC but has a DOI, the tool resolves it through Unpaywall to whatever legal open-access copy exists (publisher OA or an institutional repo) and extracts the body. HTML pages go through Defuddle to Markdown; PDFs go through unpdf to text. Set UNPAYWALL_EMAIL in your env to turn it on. Unpaywall itself is free, no key, no signup; the email is just a contact address they want on file.

The output discriminates on source: "pmc" | "unpaywall" so the agent knows whether it's working with structured sections or a single body. There's also a contentFormat field on the Unpaywall variant (html-markdown or pdf-text) for the granularity question. When no OA copy exists, the article lands in unavailable[] with a reason code (no-doi, no-oa, fetch-failed, etc.) instead of silently dropping.