Search PubMed biomedical literature, get abstracts, find related papers, and track citations via @pipeworx/mcp-pubmed — 5 tools, credential-free NCBI E-utilities

@pipeworx/mcp-pubmed — verified recipe via local npm install

Package: @pipeworx/mcp-pubmed (latest) on npm Transport: Library-style MCP export (not stdio) — import default export, call callTool(name, args) Auth: None (NCBI E-utilities are credential-free) Install: npm install @pipeworx/mcp-pubmed Runtime: node --experimental-strip-types (TypeScript source only, must copy out of node_modules)

Setup

npm install --prefix /tmp/pubmed @pipeworx/mcp-pubmed
cp /tmp/pubmed/node_modules/@pipeworx/mcp-pubmed/src/index.ts /tmp/pubmed/pubmed.ts
cd /tmp/pubmed
node --experimental-strip-types --input-type=module <<'EOF'
import pack from "./pubmed.ts";
const result = await pack.callTool("search_pubmed", { query: "CRISPR cancer therapy", limit: 5 });
console.log(JSON.stringify(result, null, 2));
EOF

⚠️ `node --experimental-strip-types` cannot import .ts from node_modules — copy source file to a non-node_modules path first.

5 tools

Verified calls (8 calls, 7 OK + 1 correct rejection)

search_pubmed — {query: "CRISPR cancer therapy 2024", limit: 5} → 1,103 total results, 5 PMIDs returned, query auto-expanded to MeSH terms. Latency: 1095ms.

search_pubmed (MeSH) — {query: "COVID-19[MeSH] vaccine efficacy", limit: 3} → 5,529 total results, 3 PMIDs. MeSH qualifier correctly parsed. Latency: 228ms.

search_pubmed (author) — {query: "Doudna JA[Author] CRISPR", limit: 3} → 224 total results, 3 PMIDs. Author field qualifier works perfectly. Latency: 224ms.

get_summary — {ids: "42317600,42290533,42258134"} → 3 articles with full metadata: titles, 6-14 authors each, journal names, pub_date, volume/issue/pages, DOI, PubMed URLs. Batch mode works correctly. Latency: 424ms.

get_abstract — {id: "42317600"} → Full unstructured abstract (DLBCL autologous transplant study), 1,200+ chars. Includes title. Latency: 568ms.

get_related_articles — {pmid: "23287722", limit: 5} (the landmark 2013 Doudna/Charpentier CRISPR-Cas9 paper) → 5 related papers, all CRISPR genome editing papers (2013-2014), relevance-ranked. Includes the companion Cong et al. paper (23287718) and Yang et al. methods paper. Latency: 919ms.

get_citations — {pmid: "23287722", limit: 5} → 4,931 citing papers in PubMed Central graph (a floor — total real citations higher). Top 5 most recent: gene editing for hearing loss (2026), keratinocyte CRISPR (2026), chromatin off-target effects (2026). Latency: 1177ms.

get_abstract (nonexistent) — {id: "999999999999"} → Correct rejection: "No abstract found for PubMed ID: 999999999999". No crash.

Key observations

1. `get_related_articles` and `get_citations` are the killer features — not available in mcp-simple-pubmed (the other PubMed MCP server). Related articles use NIH's computed neighbors (shared terms/MeSH/citations); citations are forward-citation search from PMC.
2. `query_translation` field is invaluable — shows exactly how PubMed expanded your query (e.g., "CRISPR" → MeSH terms + all-fields OR).
3. Batch `get_summary` — much cheaper than per-ID; up to ~200 IDs per call.
4. Built-in retry with backoff on 5xx/timeou