convez.ai
MCP ServerBusca & PesquisaIntermediário
kak4343

Scholar MCP – Busca Unificada em PubMed, arXiv e Semantic Scholar

kak4343/scholar-mcp

Busque 3 bases acadêmicas de uma vez, exporte para Notion instantly

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que integra três grandes bases de dados acadêmicas (PubMed, arXiv, Semantic Scholar) em uma única ferramenta com resultado unificado. Inclui cache em dois níveis para otimizar consultas repetidas e exportação direta para Notion, eliminando a necessidade de chamar múltiplas APIs.

Instalação

$npx convez-ai kak4343/scholar-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Revisão de literatura cross-disciplinar em clínica e pesquisa
  • Agregar artigos de múltiplas fontes com schema consistente
  • Organizar resultados de pesquisa diretamente em banco de dados Notion

Veredito

Ferramenta prática que reduz fricção em pesquisa acadêmica ao unificar três fontes e oferecer cache inteligente. Implementação robusta com suporte a Notion e tratamento de erros granular, útil especialmente para pesquisadores que precisam consultar múltiplas bases regularmente.

README

A tradução para o português deste README ainda está na fila.

scholar-mcp

Unified MCP server for PubMed + arXiv + Semantic Scholar scholarly search, with two-tier caching and one-call Notion export.

Search three major scholarly databases from a single MCP tool call, returning unified results (title, authors, abstract, DOI, publication date, citation count, venue, URL). Push any result set straight into a Notion database with a second call.

Why?

Existing MCP servers cover a single source (PubMed-only, arXiv-only). For clinicians and researchers doing cross-disciplinary literature reviews, calling three separate tools is friction. scholar-mcp is one tool, three sources, unified schema, with caching that makes follow-up queries instant.

Install

npm install -g @kak4343/scholar-mcp

Or run via npx without install:

npx @kak4343/scholar-mcp

Note: the package is scoped (@kak4343/) because the unscoped name scholar-mcp is already taken on npm by an unrelated package. If you see the wrong tool, double-check the scope.

Configure in Claude Desktop / Claude Code

{
  "mcpServers": {
    "scholar": {
      "command": "npx",
      "args": ["@kak4343/scholar-mcp"],
      "env": {
        "PUBMED_API_KEY": "optional, raises NCBI rate limit",
        "SEMANTIC_SCHOLAR_API_KEY": "optional, raises rate limit",
        "NOTION_TOKEN": "optional, required only for scholar_export_to_notion"
      }
    }
  }
}

Tools

scholar_search

{
  query: string;                    // e.g. "diabetic retinopathy SGLT2"
  sources?: ("pubmed" | "arxiv" | "semantic_scholar")[];  // default: all
  max_results?: number;             // default 10 per source, max 50
  date_from?: string;               // YYYY-MM-DD
  date_to?: string;                 // YYYY-MM-DD
}

Returns a unified JSON object with results array containing each paper's source, title, authors, abstract, DOI/PMID/arXiv ID, publication date, venue, citation count (Semantic Scholar only), and URL. Also returns cache_hit (boolean) and cache_hit_rate (0.0-1.0).

scholar_export_to_notion (v0.2)

Push a result set into a Notion database in one call.

{
  results: SearchResult[];           // output of scholar_search
  notion_database_id: string;        // UUID, with or without dashes
  notion_token?: string;             // overrides NOTION_TOKEN env
  include_japanese_summary?: boolean;  // default false
  dry_run?: boolean;                 // default false, build payload only
}

Each Notion page is created with these properties (the database must define them):

  • Title (title)
  • Authors (rich_text)
  • DOI (url, normalised to https://doi.org/...)
  • Source (select: pubmed / arxiv / semantic_scholar)
  • Published (date)
  • Venue (rich_text)
  • Citation Count (number)
  • Abstract (rich_text, truncated at 2000 chars; the full abstract is also written as body blocks)
  • Japanese Summary (rich_text, only when include_japanese_summary: true and the field exists)
  • Status (select: defaults to To Read)

Errors are captured per result so a single bad row never aborts the batch.

Caching (v0.2)

scholar_search is backed by a two-tier cache so repeated queries are free.

  • Memory: lru-cache, 1000 entries, 1-hour TTL
  • Disk: SQLite (better-sqlite3) at ~/.scholar-mcp/cache/scholar_cache.db, 7-day TTL
  • Key: sha256(JSON.stringify({tool, query, sources, date_from, date_to, max_results})) — query is lower-cased and sources are sorted, so semantically identical requests share a cache entry
  • Promotion: disk hits are promoted back into memory automatically
  • Reporting: every response includes cache_hit_rate (lifetime hit rate of the running process)

To wipe the on-disk cache: rm -rf ~/.scholar-mcp/cache/.

Notion configuration

scholar_export_to_notion looks up the Notion integration token in this order:

  1. notion_token argument to the tool call
  2. NOTION_TOKEN environment variable
  3. ~/.scholar-mcp/config.json — { "notion_token": "secret_..." }

Your Notion integration must be added to the target database (the integration's "Connections" must include that page). The schema above is intentionally compatible with the maintainer's personal Knowledge database (data_source_id: 0a489d15-83e8-471d-ba1e-f04030473967), so the same Notion DB can be populated by either scholar-mcp or by Python pipelines built around it.

API keys

Set them via the environment variables PUBMED_API_KEY, SEMANTIC_SCHOLAR_API_KEY, and NOTION_TOKEN.

Development

git clone https://github.com/kak4343/scholar-mcp
cd scholar-mcp
npm install
npm run build
npm start

Quick smoke test (cache hit/miss + Notion dry-run schema):

node scripts/smoke_test.mjs

Roadmap

  • v0.1 (internal): scholar_search unified across 3 sources
  • v0.2 (internal): two-tier cache (LRU memory + SQLite disk), scholar_export_to_notion
  • v0.3.0 (current, first public npm release): @kak4343/ scope rename, USER_AGENT/clone-URL fixes, README polish
  • v0.4: dedup precision (preprint ↔ published linking), Japanese summary prompt tuning for clinical triage
  • v0.5: scholar_get_related tool (Semantic Scholar citation graph)
  • v0.6: full-text retrieval for open-access papers

Legal / ethical notes

  • Only publicly indexed paper metadata and abstracts are retrieved (full text is open-access only and reserved for v0.5).
  • Cite the original source in any downstream use; every result includes a canonical URL.
  • Do not paste patient-identifying information into queries — the maintainer's project policy forbids sending such data to any cloud service, and the same applies here.

License

MIT

Author

Built by a physician using Claude Code for daily literature review. Issues and PRs welcome.

Ver o original no GitHub

Bom para

Relacionados