trove-mcp

Give Claude the ability to search, read, and synthesize academic research across 250M+ papers using free/open sources.

What it connects to

• OpenAlex (primary metadata, discovery, citations)
• Semantic Scholar (semantic recommendations and similarity)
• arXiv (preprint metadata and full text)
• Unpaywall (OA discovery by DOI)
• PubMed (biomedical indexing)
• Hugging Face Papers API (discovery context endpoints; strict trending uses citation snapshots only)
• CORE (full-text fallback)

Install (Claude Desktop, stdio)

No local install is required with npx.

Quick smoke test:

npx -y trove-mcp@latest sync --queries="graph neural network"

Optional global install:

npm i -g trove-mcp
trove-mcp sync --queries="graph neural network"

Claude Desktop on macOS

1. Open Claude Desktop.
2. Go to Settings -> Developer -> Edit Config.
3. Edit ~/Library/Application Support/Claude/claude_desktop_config.json.
4. Add this server entry under mcpServers:

{
  "mcpServers": {
    "trove": {
      "command": "npx",
      "args": ["-y", "trove-mcp"],
      "env": {
        "TROVE_CONTACT_EMAIL": "[email protected]",
        "TROVE_DB_PATH": "/Users/you/.trove-mcp/trove.db",
        "SEMANTIC_SCHOLAR_API_KEY": "",
        "UNPAYWALL_EMAIL": "[email protected]",
        "CORE_API_KEY": ""
      }
    }
  }
}

5. Save the file and fully restart Claude Desktop.
6. Start a new chat and use Trove tools.

No API keys are required to start. Missing optional keys produce graceful partial/degraded responses.

Provider hardening behavior

• Every tool returns a structured envelope with status, degraded, and warnings.
• Every tool envelope also includes meta.version so clients can verify the running package version.
• Provider outages/rate limits are reported as warnings (not process crashes), and other sources are still used.
• Semantic Scholar is publicly accessible without a key, but unauthenticated traffic is shared and may return 429; Trove reports this as degraded and falls back to OpenAlex where possible.
• Semantic Scholar recommendations are effectively API-key-backed for reliable use. find_similar_papers now fails fast with a clear warning when SEMANTIC_SCHOLAR_API_KEY is not configured.
• Unpaywall is only used when UNPAYWALL_EMAIL (or TROVE_CONTACT_EMAIL) is set.
• CORE works without key for basic access but may rate-limit heavily; set CORE_API_KEY for higher/stabler throughput.
• get_trending_papers is strict fail-closed: only mode = snapshot when non-zero citation velocity evidence exists; otherwise mode = unavailable.
• get_trending_papers self-bootstraps local citation snapshots from OpenAlex, Semantic Scholar, and Hugging Face candidate discovery, then reports snapshot_coverage diagnostics per source.
• get_trending_papers requires local snapshot history to become useful. Official provider APIs do not expose day-level citation history, so first-run mode = unavailable is expected until at least one older snapshot date exists.
• find_similar_papers is strict fail-closed: Semantic Scholar only; no lexical fallback output.
• compare_papers and build_literature_map return explicit insufficiency errors when extraction evidence quality is too low.
• PapersWithCode runtime enrichment is disabled in strict mode.
• get_references is strict fail-closed. Some large institutional arXiv preprints are retrievable as papers but still lack usable reference coverage in both Semantic Scholar and OpenAlex; in those cases Trove returns an explicit warning instead of guessed references. Use get_full_text to inspect inline citations directly when this happens.
• get_author returns mostCitedPaperIds and recentPaperIds as best-effort enrichment. If those follow-up lists cannot be fetched reliably, Trove returns the author profile as partial and does not treat empty arrays as authoritative cache data.
• OpenAlex-heavy search can still be domain-ambiguous for niche queries; search_papers and build_literature_map prefer precision gates, but queries with overloaded terms may still need tighter prompts or filters.
• trace_idea uses heuristic origin candidate ranking; timeline quality is often good, but the earliest canonical paper can still be missed when provider search ranking is imperfect.

HTTP mode (streamable)

TROVE_HTTP_BEARER_TOKEN=change-me npx trove-mcp --transport=http --port=3000

Browser client example:

TROVE_HTTP_BEARER_TOKEN=change-me TROVE_HTTP_CORS_ORIGIN=http://localhost:3000 npx trove-mcp --transport=http --port=3000

• POST /mcp for authenticated MCP requests
• GET /health for health checks
• Auth header: Authorization: Bearer <token>
• CORS exposes Mcp-Session-Id/MCP-Session-Id for browser MCP clients

Tools

| Tool | What it does | |---|---| | search_papers | Multi-source search + dedupe + deterministic ranking (OpenAlex/S2/arXiv/PubMed/CORE) | | get_trending_papers | Topic papers ranked by citation velocity with mode = snapshot | unavailable | | get_paper | Resolve paper by DOI/arXiv/S2/OpenAlex/PubMed/title | | get_full_text | arXiv -> Unpaywall -> CORE full-text fallback with chunked output | | get_citations | Papers that cite a target paper | | get_references | Papers referenced by a target paper | | find_similar_papers | Semantic Scholar recommendations; reliable use requires SEMANTIC_SCHOLAR_API_KEY | | get_author | Author profile and impact metrics; paper-list enrichment is best-effort | | get_institution_output | Institution profile + publication output | | get_coauthor_network | Collaboration graph around an author | | build_literature_map | Structured evidence map (claims/methods/limitations/consensus) | | compare_papers | Structured 2-5 paper comparison | | trace_idea | Concept lineage across time and influence |

Resources

• trove://resources/version
• trove://resources/source-capability-matrix
• trove://resources/schema-reference
• trove://resources/cache-health

Prompts

• literature-review-workflow
• paper-comparison-workflow
• idea-lineage-workflow

Sync job for citation snapshots

npx trove-mcp sync
npx trove-mcp sync --queries="agentic ai,graph neural network,causal inference"

Run this on a schedule (e.g. cron) to improve get_trending_papers quality. Without older local snapshots, first-run trending results will correctly return mode = unavailable.

Development

npm install
npm run typecheck
npm test
npm run test:quality
npm run verify:release
npm run build
npm run inspect

Live-contract tests:

LIVE_CONTRACT=1 npm run test:live

Temporarily hide tools from MCP registration (for strict release gating):

TROVE_DISABLED_TOOLS=get_trending_papers,find_similar_papers npx -y trove-mcp@latest

Data and compliance

• No scraping
• No Sci-Hub or paywalled bypassing
• All sources are public/open APIs or legal OA links

License

MIT