A research librarian that never sleeps, for your LLM

What it does

paper-search-mcp plugs into Claude Desktop or other MCP clients to search, download, and extract text from academic papers across arXiv, PubMed, bioRxiv, Semantic Scholar, and roughly fifteen other sources. It standardizes everything into a consistent Paper format so your LLM doesn’t have to parse twenty different JSON schemas.

The interesting bit

The project is admirably honest about what breaks. A live-tested capability matrix shows exactly which platforms support search, download, or full-text read — Google Scholar is ⚠️ for bot detection, PubMed is search-only, Sci-Hub is a fallback of last resort. The README even documents upstream failure modes (CORE 500s, OpenAIRE 403s, CiteSeerX archive redirects) rather than pretending they don’t happen. That’s unusual infrastructure transparency.

Key highlights

• Two-layer architecture: unified search_papers with deduplication, plus modular per-platform connectors with DOI extraction
• Free-first design: public APIs are default; optional keys only for rate limits or stability
• OA-first fallback chain: native download → repository discovery → Unpaywall DOI resolution → optional Sci-Hub
• Graceful degradation: connectors retry with backoff, escalate request profiles, or return empty rather than crash
• IEEE Xplore and ACM DL included as disabled-by-default skeletons — no surprise API calls

Caveats

• Google Scholar and SSRN are effectively broken without proxy configuration due to bot detection
• BASE requires institutional IP registration; returns empty otherwise
• Unpaywall is entirely disabled without an email environment variable
• IEEE and ACM connectors are skeleton-only, not functional integrations

Verdict

Worth it if you want an LLM assistant that can actually retrieve and cite papers without making up sources. Skip if you need reliable full-text access to paywalled content — this project deliberately avoids that lane.