paper-search-mcp

Academic paper search MCP server for Cloudflare Workers, designed for deployment via mcp-deploy.

Searches across multiple platforms with unified ranking via Reciprocal Rank Fusion (RRF):

• Semantic Scholar — CS, social science, interdisciplinary; ML relevance ranking, TLDRs, influential citations, bulk Boolean search, paper recommendations

• OpenAlex — 250M+ works; semantic search via GTE-Large embeddings, journal/topic/OA filtering

• CrossRef — 150M+ records; DOI lookups, field-specific queries, journal metadata

• arXiv — CS, math, physics preprints; query syntax (ti:, au:, abs:, cat:)

• PubMed — Biomedical literature via NCBI E-utilities; MeSH terms

• bioRxiv / medRxiv — Biology and medical preprints; category filtering

Core platforms (always on): Semantic Scholar, CrossRef, OpenAlex — provide quality scoring signals (FWCI, h-index, venue quality, citations). Optional platforms (configurable via ENABLED_PLATFORMS): arXiv, PubMed, bioRxiv, medRxiv. Default: arxiv.

Tools

Search

Lookup & Discovery

Key features

RRF Fusion — search_papers, search_recent, and search_journal merge results from multiple platforms using Reciprocal Rank Fusion (score = Σ 1/(k + rank), k=60). Papers are deduplicated by DOI with metadata merged from the richest source.

Semantic Search — OpenAlex semantic search uses GTE-Large embeddings over 217M works to find conceptually related papers even with different vocabulary. Enable with semantic=true. Requires OPENALEX_API_KEY.

Paper Recommendations — find_similar_papers uses Semantic Scholar's ML engine. Provide positive seed paper IDs (and optional negative IDs to steer away from) to get ranked recommendations.

Bulk Boolean Search — Semantic Scholar bulk search supports +required, -excluded, "exact phrase", and | OR operators. Enable with bulk=true for high-recall retrieval up to 1000 results.

Date Filtering — search_papers supports date_from and date_to (YYYY-MM-DD) to scope results to a time window. Dates are mapped to platform-native filters where possible, with post-fusion filtering as a safety net.

Sort & Quality Control — sort_by (relevance/date/citations) controls result ordering. min_citations filters low-quality or uncited results (with a warning when preprint sources are heavily filtered).

How it works

This repo contains only MCP logic. Auth, deployment, and UI are handled by mcp-deploy. The repo ships:

• src/ — MCP server code (Cloudflare Workers + Durable Objects)

• mcp-deploy.json — deployment contract (secrets, config, worker settings)

All platforms return a normalized Paper interface with consistent fields (title, authors, abstract, doi, url, pdf_url, published_date, citations, etc.). Platform-specific data is stored in the extra field.

Configuration

*OpenAlex keyword search works without a key; semantic search and higher rate limits require one.

Local development

npm install
npx wrangler dev
# Health check: http://localhost:8787/

Testing

npm test

Integration tests hit live APIs. Set .dev.vars with optional API keys for better rate limits:

SEMANTIC_SCHOLAR_API_KEY=...
PUBMED_API_KEY=...
OPENALEX_API_KEY=...
CONTACT_EMAIL=you@example.com

Release

Tag a version to trigger the GitHub Actions release workflow:

git tag v0.2.3
git push --tags

This builds worker.mjs and publishes it alongside mcp-deploy.json as release assets. mcp-deploy fetches these assets to deploy the worker.