Pubtator

What the Pubtator MCP server does

The Pubtator MCP server connects MCP clients to the PubTator 3 API operated by NCBI and NLM. It searches biomedical literature and exposes machine-extracted links among genes, diseases, chemicals, and variants. The underlying corpus covers roughly 36 million PubMed abstracts and PMC open-access full-text records.

Entity lookup resolves names such as BRCA1 or Breast Neoplasms to PubTator’s name-based identifiers. Literature searches can use free text, entity IDs combined with AND or OR, or relation expressions such as a chemical that treats a disease. Results include publication identifiers and bibliographic fields, along with the sentence matched by the query and decoded entity mentions.

This fits the Biology, Medicine and Bioinformatics category when an agent needs to find biomedical papers, inspect entity mentions, or explore model-extracted gene–disease–chemical–variant relationships.

How it works

The Pubtator MCP server provides five focused operations. pubtator_find_entity uses autocomplete to resolve a name and can filter by gene, disease, chemical, variant, species, or cell line. pubtator_search returns ten results per page and reports totals and page counts. pubtator_relations ranks extracted relationships by supporting publication count and can include supporting passages for the highest-ranked results.

For auditing one relationship, pubtator_relation_evidence returns all available supporting passages in pages of ten. pubtator_annotations accepts up to 20 PMIDs and returns normalized mentions with character offsets, plus relations detected within each article. Setting full_text enables PMC full-text annotations instead of title-and-abstract coverage.

Entity arguments that are not already in @TYPE_Name form are resolved through autocomplete. PubTator identifiers are case-sensitive and use names rather than backing database IDs. The API supports twelve relation types, including associate, cause, cotreat, inhibit, interact, prevent, stimulate, and treat.

Setup and configuration

The hosted endpoint can be added to an MCP client with the URL https://gateway.pipeworx.io/pubtator/mcp. The README specifically shows configurations for Claude Desktop, Cursor, and Windsurf. The same gateway also offers a broader endpoint at https://gateway.pipeworx.io/mcp; either endpoint can access gateway routing through ask_pipeworx, while the scoped endpoint lists this pack’s tools directly alongside shared gateway tools.

For a local connection, run npx -y @pipeworx/mcp-pubtator as a stdio MCP server. This standalone mode does not require a Pipeworx account and exposes only the Pubtator pack’s tools. The service is keyless. It can also be called over HTTP through the gateway’s tool endpoint, such as POST /v1/tools/pubtator_find_entity.

Limitations and notes

Relations and annotations are predictions from PubTator 3’s extraction models, not human-curated facts. The supporting_publications value measures model evidence count and should not be treated as a curation or validation status. Supporting passages are useful for reviewing how a relation was extracted.

The upstream page size is fixed at ten, and there is no size parameter. Full-text requests may produce more than 100 passages; passages can be truncated at 1,500 characters. Common species or cell-line names may return no result with a concept filter, so omitting that filter can be necessary. Relation searches do not depend on the order of the two entities.