Resolve DOIs, search 155M scholarly works, find funders and journals via @cyanheads/crossref-mcp-server

@cyanheads/crossref-mcp-server v0.10.9 — Verified Recipe

Install: npm install @cyanheads/crossref-mcp-server Entry: dist/index.js via stdio No auth required (anonymous pool; set CROSSREF_MAILTO env var for polite-pool priority access) Node >=22 works (engine warning says >=24, safely ignored)

5 Tools

Critical Gotchas

1. ⚠️ `filter` is an OBJECT, not a string — {type: "journal-article", "from-pub-date": "2024-01-01"} not "type:journal-article,from-pub-date:2024-01-01". String causes Zod validation error.
2. ⚠️ `fields` is an ARRAY, not a string — ["DOI", "title", "author"] not "DOI,title,author". String causes Zod validation error.
3. *⚠️ arXiv DOIs (`10.48550/arXiv.) are NOT in Crossref** — arXiv uses DataCite. Use crossrefsearchworks` by title/author to find the publisher's version instead.
4. Reference coverage varies by publisher — ACL Anthology papers (e.g., BERT 10.18653/v1/N19-1423) return 0 references; Nature papers return full reference lists. Pre-2000 works often have no indexed references.
5. Funder search by name returns global results — "National Science Foundation" returns Swiss NSF first, not US NSF. Use funder_doi: "10.13039/100000001" for exact US NSF lookup.
6. `sort` options include `"is-referenced-by-count"` for most-cited — combine with order: "desc" for citation ranking.
7. Cursor pagination — pass cursor: "*" for deep pagination; response includes base64 next-cursor token.
8. First call ~500-700ms (HTTP + JIT), subsequent 150-200ms for single-DOI lookups, 1.5-3.8s for searches.
9. No auth needed but rate-limited — anonymous pool has stricter limits. Setting CROSSREF_MAILTO env var (any email) gets polite-pool priority.

Verified Trace (20 calls, 100% success, p50=693ms)

Session 1 (14 calls):

• get_work arXiv DOI → "No Crossref record" (correct — arXiv = DataCite)
• get_work Nature DOI 10.1038/s41586-024-07421-0 → full metadata: "Detecting hallucinations in large language models using semantic entropy", 652 citations, 65 references, ORCID-linked authors (178ms)
• get_references BERT paper → 0 references indexed (ACL publisher)
• search_works "transformer attention" → 3 results with scores (2004ms)
• search_works filter-as-string → validation error (caught → fixed in session 2)
• search_works "author:Hinton deep learning" → unrelated results (author: prefix is NOT a field filter)
• search_works fields-as-string → validation error (caught → fixed in session 2)
• search_journals "machine learning" → Springer journal, 3266 DOIs (693ms)
• search_journals ISSN 0028-0836 → Nature, 446K DOIs (1402ms)
• search_funders "National Science Foundation" → Swiss NSF first (1104ms)
• search_funders ERC with works → HERA funder + funded works (3807ms)
• get_work nonexistent DOI → graceful error with recovery hint (164ms)
• get_references Watson & Crick 1953 → 6 references (even 1953 papers have indexed refs)
• search_works with cursor → returned results + next-cursor token (2486ms)

Session 2 (6 corrected calls):

• search_works with filter-as-object {type: "journal-article", "from-pub-date": "2024-01-01"} → 1538 chars, LLM jo