pubmed-mcp-server

Public Hosted Server: https://pubmed.caseyjhand.com/mcp

Overview

Biomedical literature from PubMed, PubMed Central, and Europe PMC. Search it, fetch metadata and full text, resolve identifiers and partial citations, format references, and ground queries in MeSH vocabulary. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above. Ships the pubmed-guide agent skill (skill://pubmed-guide/SKILL.md), which MCP clients read for PubMed query syntax, DOI-to-PMID routes outside PMC, reference-list resolution, citation counts, and full-text misses.

Tools

Resources

Prompts

Related MCP server: PubMed MCP Server

Capability reference

pubmed_search_articles tool

• Full PubMed query syntax plus filters (author, journal, MeSH, publication type, language, species, abstract, free full text) and publication/modification/Entrez date ranges

• query is optional: with an author, journal, MeSH, or publication-type filter, the filters alone make the search (a journal's output over a date range, for example)

• Up to 1,000 results per page and offsets up to 9,998, or maxResults: 0 for the match count alone; summaryCount adds briefs for up to 50 hits. A larger page size is served at the cap with a notice naming where to continue; a page that reaches PubMed's 9,999-record window says how many matches no offset can reach and advises narrowing or splitting the query by dateRange

• Reports totalCount, the effectiveQuery PubMed ran, and the appliedFilters

• Rejects blank filter values, impossible dates, and reversed date ranges before PubMed is called

pubmed_fetch_articles tool

• Up to 200 PMIDs per call; misses are listed in unavailablePmids

• Abstract, authors, journal, DOI, publication types, linked retraction/erratum/comment notices, and links; MeSH terms on by default, grants via includeGrants, and the members of a collective author (its InvestigatorList, thousands on some consortium trials) via includeInvestigators

• A retracted, corrected, or questioned article opens with the same notice block as pubmed_format_citations, directly under its title

• dataBanks lists the record's trial registrations and data deposits (ClinicalTrials.gov NCT IDs, GenBank and PDB accessions, Dryad DOIs) verbatim; search pubmed_search_articles for <accession>[si] to find other records citing one

• recordType separates journal articles from Bookshelf chapters and books; opt-in maxResponseCharacters defers overflow to deferred.ids

pubmed_fetch_fulltext tool

• pmcids, pmids, and dois in any mix, up to 50 distinct identifiers per call: the first 10 are fetched and the rest returned in overLimit for a follow-up call. Each is tried against PMC, then Europe PMC, then Unpaywall (needs UNPAYWALL_EMAIL); an article several of them name is fetched and returned once, and viaSource names the tier that answered

• source: "pmc" returns sections (the body, then back matter such as acknowledgments, declarations, and appendices), tables, and figures; source: "unpaywall" returns a best-effort HTML or PDF text body

• Europe PMC full text is fetched for records Europe PMC marks open access, preprints included; articles carry the licenses their PMC JATS states

• Misses carry a typed reason and triedTiers; sections, maxSections, maxCharacters, overflowMode, and maxResponseCharacters bound the response, truncation reports what was cut, and sectionCap names the sections maxSections cut

pubmed_europepmc_search tool

• Adds preprints (PPR), patents (PAT), and Agricola (AGR) to MED and PMC; defaults to MED, PMC, PPR

• Cursor pagination via cursorMark / nextCursorMark, up to 100 per page; a larger page size is served at the cap with a notice naming where to continue. Abstracts arrive as 400-character snippets

• Optional sort of comma-separated <field> asc|desc keys; relevance is the default order, and a sort Europe PMC cannot use fails as europepmc_invalid_input

• Hits carry license and hasFullTextXml

• Not registered when EUROPEPMC_ENABLED=false

pubmed_europepmc_fetch tool

• Up to 25 records per call as records: [{ source, epmcId }], each pair copied from a search hit, with the full abstract, license, and hasFullTextXml; unresolved ids land in notFound; opt-in maxResponseCharacters defers overflow to deferred.records

• Not registered when EUROPEPMC_ENABLED=false

pubmed_fetch_citation_metrics tool

• Up to 1,000 PMIDs per call; the first 200 distinct are fetched and the rest returned in overLimit, PMIDs NIH iCite has no record for in unavailablePmids

• Relative Citation Ratio (with a provisional flag), NIH percentile, citation count and rates, Approximate Potential to Translate, MeSH category shares, and clinical citations

• iCite does not flag retractions; check pubmed_fetch_articles for notices

• Source: NIH iCite

pubmed_format_citations tool

• Up to 50 PMIDs per call, in any mix of apa, mla, bibtex, ris, and vancouver

• Bookshelf records cite as edited books, each chapter dated by its own last revision or contribution date; PMIDs that can't be fetched are reported as unavailable

• Vancouver writes page ranges in elided Citing Medicine form (1637-1645 → 1637-45); BibTeX issn and RIS SN carry the journal's linking ISSN, and RIS TI drops its trailing period

• Retractions, errata, and expressions of concern NCBI links to an article come back as notices and are shown ahead of its citations; Vancouver adds a Retraction in: / Erratum in: note and APA a retraction parenthetical

pubmed_find_related tool

• similar, cited_by, or references, up to 50 per page with offset pagination. A larger page size is served at the cap with a notice naming where to continue

• Rows carry the same identifiers as search summaries: doi, pmcId and pmcUrl when the record has them, and a pubmedUrl on every row

• Falls back to Europe PMC, then OpenAlex, and names the provider that answered; fails as all_providers_failed if none can

pubmed_spell_check tool

• Returns original, corrected, and hasSuggestion; corrections land in place, so field tags, operators, spacing, and gene symbols stay as written, and a single letter such as an author initial is never swapped for another

• Values of indexed field tags (aged[mh], zhang y[au], free full text[sb]) are kept as written; free-text fields ([tiab], [ti], [tw]) are still corrected

• Run it after a zero-hit or thin search, then retry pubmed_search_articles with corrected

pubmed_lookup_mesh tool

• MeSH records of every type (descriptors, supplementary concepts, qualifiers, pharmacological actions) by name or free-text term, with the exact heading match pinned first: the record the query names, a descriptor before a qualifier or pharmacological-action record of the same name, or the heading an entry term maps to; up to 50 per page, continued via nextOffset. A larger page size is served at the cap with a notice naming where to continue

• Records carry their MeSH unique ID (meshId, e.g. D000072657) beside the Entrez UID, their recordType, and a searchTerm, the PubMed clause that searches them ("Name"[mh], [nm] for a supplementary concept, [pa], [pt], [sh]); a drug-class heading returns its descriptor and its pharmacological-action record as two records under one name and ID, told apart by recordType

• includeDetails (default on) adds tree numbers, scope notes, entry terms, yearIntroduced, and previousIndexing; the exact heading match on a query's first page also gets its allowable subheadings

pubmed_lookup_citation tool

• Up to 25 partial citations per call; a journal or a four-digit year is required, and volume, first page, and author sharpen the match

• Each is matched independently and comes back matched, not_found, or ambiguous, with recovery detail

pubmed_convert_ids tool

• Up to 50 ids per call, all of one declared idType (doi, pmid, pmcid); only PMC-indexed articles resolve

• One success or error row per id, in order, so a partial batch never fails

pubmed://database/info resource

• Live EInfo call for the pubmed database, returned as application/json

• count, lastUpdate, and fields[], whose names are the search tags pubmed_search_articles accepts

research_plan prompt

• Arguments: title, goal, and keywords required; organism and includeAgentPrompts optional

• Returns a four-phase research plan; includeAgentPrompts: "true" adds an agent-guidance block under each of its nine sub-steps, two of which point at tools: the literature review names pubmed_search_articles and pubmed_lookup_mesh, and the interpretation step names pubmed_search_articles

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

PubMed-specific:

• NCBI E-utilities (ESearch, EFetch, ESummary, ELink, ESpell, EInfo, ECitMatch) and the PMC ID Converter, with Europe PMC, OpenAlex, and Unpaywall filling the gaps

• Shared NCBI request queue: paced request starts, capped concurrency, a cooldown that holds every caller after an NCBI 429, one deadline covering queue wait and retries, and fail-fast for an endpoint whose retries were just spent on an outage: later calls to it make one attempt each until one succeeds

• XML parsing built for PubMed's inconsistent records: structured abstracts, missing fields, varying date formats

• Forgiving identifiers: a zero-padded PMID resolves as the PMID it spells in every tool that takes one; pubmed_fetch_fulltext also matches DOIs case-insensitively and PMC IDs with or without the PMC prefix

• Hand-rolled citation formatters (APA, MLA, BibTeX, RIS, Vancouver) with no dependencies

Agent-friendly output:

• Provenance in responses: source labels, license fields on full-text and Europe PMC results, best-effort warnings on Unpaywall results, and effective-query echo on searches, so agents can judge what to trust

• Graceful partial failure: batch tools return per-item success/error rows instead of failing the request, with structured status codes and actionable next-step text

• Discriminated output contracts: source: "pmc" | "unpaywall", typed unavailable reasons, viaSource and triedTiers fields, so callers branch on data, not string parsing

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http pubmed-mcp-server https://pubmed.caseyjhand.com/mcp

Or run it locally:

claude mcp add pubmed-mcp-server -- npx -y @cyanheads/pubmed-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "pubmed-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/pubmed-mcp-server"]
    }
  }
}

An NCBI API key in NCBI_API_KEY raises the rate limit from 3 to 10 requests per second (-e NCBI_API_KEY=your-key, or an env entry). For Docker or a source build, see Running the server.

Configuration

See .env.example for every server setting and the common framework overrides.

Running the server

Local development

Requires Bun v1.4.4 or higher (or Node.js v24+).

• Clone and install:

  git clone https://github.com/cyanheads/pubmed-mcp-server.git
  cd pubmed-mcp-server && bun install
  cp .env.example .env   # set NCBI_API_KEY, NCBI_ADMIN_EMAIL, and UNPAYWALL_EMAIL as needed

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite

Docker

docker run --rm -p 3010:3010 ghcr.io/cyanheads/pubmed-mcp-server:latest

The image defaults to HTTP transport, stateless session mode, and logs to /var/log/pubmed-mcp-server. To run it over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/pubmed-mcp-server:latest"]. To build locally, docker build -t pubmed-mcp-server .; OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for logging, ctx.state for storage

• Register new tools and resources in the createApp() arrays in src/index.ts

• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.