Trove

npm package · GitHub repo

Trove is an MCP server for academic research.

The basic idea is simple: let Claude search papers, fetch metadata, read open full text when it can, and build structured outputs on top of that.

I have tried to keep it conservative. If a provider cannot support an answer well enough, Trove returns partial, error, or unavailable instead of pretending it knows more than it does.

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.

Using with Claude

If you want Claude to use Trove, just say so in the prompt.

Examples:

• "Use Trove to find papers on speculative decoding"
• "Use Trove to compare DPO and RLHF papers"
• "Use Trove to trace the origin of LoRA"

If you do not want Claude to touch the setup-sensitive tools, hide them with TROVE_DISABLED_TOOLS.

What to expect on a fresh install

• get_trending_papers will usually return mode = unavailable on first run. Trove has to accumulate at least two local citation snapshot dates before it can compute citation velocity honestly.
• find_similar_papers is usable without a Semantic Scholar key only in theory. In practice, reliable semantic recommendations require SEMANTIC_SCHOLAR_API_KEY, and Trove says so directly.
• Some papers are retrievable even when their provider coverage is incomplete. For example, an arXiv preprint may have full text but no usable reference list in Semantic Scholar or OpenAlex.

Runtime behavior

• Every tool returns a structured envelope with status, degraded, warnings, provenance, meta.version, and tool-specific data.
• Provider outages and rate limits are surfaced as warnings, not process crashes.
• Trove prefers explicit unavailability over approximate output. That is why some tools return empty results with warnings instead of "best effort" guesses.

Tools with real prerequisites

• get_trending_papers
  • Strict fail-closed: only returns mode = snapshot when local citation history contains at least two snapshot dates and the computed velocity is non-zero.
  • Candidate discovery can come from OpenAlex, Semantic Scholar, and Hugging Face, but none of those official APIs provide day-level historical citation data.
  • First-run mode = unavailable is expected. Run trove sync once to seed, then rerun after a later day to accumulate history.
• find_similar_papers
  • Strict fail-closed: Semantic Scholar only, no lexical fallback.
  • Reliable use requires SEMANTIC_SCHOLAR_API_KEY.
  • Without a key, Trove returns an explicit error immediately instead of spending time on unreliable unauthenticated requests.

Known provider constraints

• Semantic Scholar is publicly accessible without a key, but unauthenticated traffic uses a shared pool and often returns 429.
• Unpaywall is only used when UNPAYWALL_EMAIL or TROVE_CONTACT_EMAIL is set.
• CORE works without a key for basic access but may rate-limit or degrade under load; CORE_API_KEY improves throughput. Also the CORE API is free, you just need to signup on their site and verify your mail address.
• 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 and suggests get_full_text for inline citation inspection.
• get_author returns mostCitedPaperIds and recentPaperIds as best-effort enrichment. If those follow-up lists cannot be fetched reliably, Trove returns the profile as partial and does not cache empty arrays as authoritative.
• OpenAlex-heavy search can still be domain-ambiguous for niche queries. search_papers and build_literature_map apply precision gates, but overloaded terms may still need tighter prompts or filters.
• trace_idea uses heuristic origin ranking. The timeline is often useful, but the earliest canonical paper can still be missed when provider ranking is imperfect.

A practical setup

If you just want the useful core workflow, this is enough:

• TROVE_CONTACT_EMAIL
• UNPAYWALL_EMAIL
• TROVE_DB_PATH

If you want similar-paper recommendations to work reliably, also set:

• SEMANTIC_SCHOLAR_API_KEY

If you want trending to become useful, run:

npx -y trove-mcp@latest sync

and do that on a schedule. Trending depends on local snapshot history, so it will not be useful on day one.

If you want a more conservative setup, you can hide the tools that need extra setup:

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

That gives you the core paper-search / paper-read / compare / trace workflow without exposing the two tools that are most sensitive to setup.

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
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

These are optional workflow helpers.

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. If you want actual short-window trending, run this at least twice on different days.

Tests

The tests are mostly there to protect the parts that are easy to quietly break:

• identifier normalization
• dedupe / alias handling
• ranking and off-topic filtering
• full-text cleanup quality
• synthesis extraction quality
• fail-closed provider behavior
• release gates for fake trending / fake similarity / weak evidence
• live provider contract checks

npm run verify:release is in the publish path. Live tests are separate because they depend on upstream providers behaving that day.

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