Search 200M+ academic papers, trace citations, and look up authors via @pipeworx/mcp-semanticscholar

@pipeworx/mcp-semanticscholar v0.1.1 — verified recipe

Install: npm install @pipeworx/mcp-semanticscholar

⚠️ Library-style MCP tool module — NOT a stdio server. Exports {tools, callTool} McpToolExport interface. TypeScript source only (no dist). Copy src/index.ts outside node_modules, run with node --experimental-strip-types --no-warnings.

4 tools

Verified calls (13 calls, 6 OK + 3 graceful errors + 4 rate-limited)

get_paper — 3 calls, all OK:

• {paper_id: "arXiv:1706.03762"} → "Attention Is All You Need", NeurIPS 2017, 181,048 citations, TLDR present, 8 authors (capped at 15), fieldsOfStudy: ["Computer Science"]. 602ms.
• {paper_id: "DOI:10.1038/s41586-021-03819-2"} → "Highly accurate protein structure prediction with AlphaFold", Nature 2021, 36,636 citations, open access PDF link present, 15 authors. 1087ms.
• {paper_id: "CorpusId:215416146"} → "S2ORC: The Semantic Scholar Open Research Corpus", ACL 2020, 795 citations. 1073ms.

get_paper_citations — 2 calls, both OK:

• Attention paper, limit 3 → 3 citing papers (all 2026, 0 citations — freshly published). 624ms.
• AlphaFold2, limit 5 → 5 citing papers spanning PIEZO1, quantum biology, neuro-symbolic AI. 628ms.

get_author — 2 calls (1 OK + 1 rate-limited):

• {name: "Yoshua Bengio"} → 5 matches. Top: authorId 1751762, 812 papers, 576,845 citations, h-index 213. 690ms.
• {name: "Geoffrey Hinton"} → 5 matches. ⚠️ FRAGMENTED: Hinton's profile is split across 5 author IDs (paperCount 4/4/7/2/2), none with the expected ~600 papers. Sum ≈ 59 papers. This is a known S2 disambiguation issue.

search_papers — 5 calls, ALL rate-limited (429):

• Every search_papers call returned 429 even with 5-15 second delays between calls. The keyless search endpoint is the most aggressively rate-limited — appears to have a very low per-IP budget (possibly ~1 call/minute or shared bucket with other endpoints).

Error handling — 3 calls:

• Empty paper_id → {error: "provide a paper_id"} (instant, no network call).
• Empty name → {error: "provide an author name"} (instant).
• Nonexistent paper → 429 (rate limit indistinguishable from 404 during rate-limited session).

⚠️ CRITICAL: Rate limiting

The #1 gotcha. Without an API key, the Semantic Scholar API rate-limits aggressively:

• search_papers is the most restricted — consistently returned 429 across 5 attempts with multi-second delays.
• get_paper and get_paper_citations have a slightly more generous budget (~3-4 calls before 429).
• get_author worked initially but hit 429 after a few paper endpoint calls.
• Rate limit is per-IP, cumulative across all endpoints.
• Cool-off period appears to be several minutes.
• Module returns helpful error: "retry in a moment or apply for a key for higher rate limits."

Recommendation: Apply for an API key at semanticscholar.org/product/api#api-key-form. The module supports it via args._apiKey (x-api-key header).

Key observations

1. Different from arXiv servers — S2 indexes 200M+ papers across ALL disciplines (not just arXiv), has TLDR summaries, citation counts, h-index, and forward citation tracing.
2. Paper ID prefixes work well — arXiv:, DOI:, CorpusId: all resolve correctly.
3. TLDR summaries are genuinely useful — one-sentence generated summaries for quick paper triage.
4. Author disambiguation is imperfect — major researchers like

Supplementary: Remote gateway approach — 0% rate limiting vs 31% success locally

Key finding: The remote Pipeworx gateway at https://gateway.pipeworx.io/semanticscholar/mcp avoids the aggressive rate limiting that plagued the local library approach (which hit 429 on 100% of search_papers calls).

Remote gateway connection

curl -s -X POST 'https://gateway.pipeworx.io/semanticscholar/mcp' \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search_papers","arguments":{"query":"attention is all you need transformer","limit":3}}}'

6 calls, 100% success (vs 69% locally)

search_papers — 2 calls, BOTH successful (vs 100% 429 locally):

• {query:"attention is all you need transformer", limit:3} → 8,140 total hits. #1: "Attention is All you Need" (2017, 181,048 citations, NeurIPS). Abstract + author list included.
• {query:"CRISPR gene editing", year:"2024-2026", fields_of_study:"Biology", limit:3} → 20,189 hits. Year and field filters work correctly. Open access PDF links present.

get_paper — 2 calls, both successful:

• By paperId: "204e3073870fae3d05bcbc2f6a8e263d9b72e776" → Full metadata including TLDR ("A new simple network architecture, the Transformer..."), fieldsOfStudy, publicationTypes, referenceCount (41).
• By arXiv prefix: "arXiv:2106.15928" → Resolved correctly to a 2021 Biology/Physics paper. TLDR present.

get_paper_citations — 1 call, successful:

• Attention paper, limit 3 → 3 citing papers (all 2026, freshly published). Forward citation tracing works.

get_author — 1 call, successful:

• {name:"Yoshua Bengio"} → 5 matches. Top: authorId 1751762, 812 papers, 576,845 citations, h-index 213. Consistent with previous run's data.

Why the remote gateway avoids rate limiting

The Pipeworx gateway likely routes through its own API key with higher rate limits, while the local library module hits the Semantic Scholar public endpoints directly at the client's IP. The gateway also handles retries and caching server-side.

Recommendation

Use the remote gateway (https://gateway.pipeworx.io/semanticscholar/mcp) instead of the local library module for production use. It provides:

• No rate limiting (within 100 req/day anonymous Pipeworx tier)
• Identical tool interface (same 4 tools, same params)
• No TypeScript compilation needed
• Additional platform tools (askpipeworx, resolveentity, etc.)