Clinical MCP Server
odanree/clinical-mcp-server
PubMed + openFDA integrados ao Claude via MCP para pesquisa clínica
Ver no GitHubServidor MCP que expõe ferramentas de busca em literatura clínica (PubMed) e rótulos de medicamentos (openFDA) para Claude Desktop e qualquer cliente MCP. Integra APIs públicas do governo americano sem autenticação, custo ou dados sensíveis, permitindo consultas estruturadas sobre literatura médica e prescrições farmacêuticas.
Instalação
npx convez-ai odanree/clinical-mcp-server1 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
Casos de uso
- Buscar literatura médica primária sobre condições e tratamentos
- Consultar rótulos de medicamentos e informações de prescrição estruturadas
- Validar recomendações clínicas contra fontes oficiais atualizadas
- Analisar avisos de segurança (boxed warnings) e contraindicações de fármacos
Veredito

Projeto bem executado que preenche um nicho real: conectar Claude a fontes de verdade clínica estruturada (PubMed + openFDA) sem overhead de autenticação. Implementação técnica sólida com FastMCP e Pydantic, mas público-alvo é restrito (profissionais de saúde/pesquisa). Excelente exemplo de MCP production-ready.
README
A tradução para o português deste README ainda está na fila.
clinical-mcp-server
MCP server exposing clinical-literature + drug-label search to Claude Desktop and any other MCP client. Two tools, both backed by public US government APIs — no auth, no cost, no PHI.
MCP · FastMCP · Claude Desktop · Anthropic · NCBI PubMed E-utils · openFDA · Pydantic v2 · httpx · respx · pytest-asyncio
| Tool | What it does | Backed by |
|---|---|---|
search_pubmed_tool | PubMed query with field tags / MeSH / booleans → list of papers with PMID, authors, journal, DOI, year, and a clickable PubMed URL | NCBI E-utils |
get_drug_label_tool | Brand / generic / substance name → curated FDA label sections (indications, dosing, contraindications, warnings, boxed warning, interactions) + DailyMed SPL link | openFDA |
Use it from Claude Desktop
- Install
git clone https://github.com/odanree/clinical-mcp-server
cd clinical-mcp-server
python -m venv .venv && .venv/Scripts/activate
pip install -e ".[dev]"
- Drop these settings into your Claude Desktop config (
%APPDATA%\Claude\claude_desktop_config.jsonon Windows,~/Library/Application Support/Claude/claude_desktop_config.jsonon macOS):
{
"mcpServers": {
"clinical": {
"command": "C:\\Users\\Danh\\Documents\\Projects\\clinical-mcp-server\\.venv\\Scripts\\python.exe",
"args": ["-m", "server.main"],
"env": {
"NCBI_API_KEY": "",
"NCBI_EMAIL": "you@example.com"
}
}
}
}
- Restart Claude Desktop. The tool icon (🔧) in the input row should now list
search_pubmed_toolandget_drug_label_tool. Try:
"What does the recent literature say about metformin and cardiovascular outcomes in CKD patients?"
"Pull the FDA label for atorvastatin — focus on the contraindications and boxed warning."
Claude will call the tools as needed and cite the returned URLs.
Full setup notes (rate-limit gotchas, API keys, troubleshooting) live in docs/claude-desktop-config.md.
Use it as a remote HTTP MCP server
For custom clients or web-hosted setups:
python -m server.main --transport http --host 0.0.0.0 --port 8765
# now serving the MCP streamable-HTTP transport at http://127.0.0.1:8765/mcp
A LangGraph agent in agent/ (next session) will consume this transport directly.
Why this exists
MCP is the emerging standard for letting an LLM call external tools without bespoke per-client glue. Most senior-AI portfolios still don't have an MCP project — this is the cheapest possible "I'm tracking the bleeding edge" signal you can ship.
Why these two tools specifically:
- PubMed is the canonical source of clinical primary literature. Every clinical question worth grounding hits it.
- openFDA labels are the only source of structured prescribing information that's machine-readable and free. Without them, an LLM has to summarize from training data — which is exactly the hallucination case.
What's left for v2 (agent layer, in agent/ when it lands):
- LangGraph supervisor that consumes this MCP server as its tool provider
- 20-case golden eval set (clinical questions with known answers + sources)
- Compliance workflow: "given guideline X, find all primary studies it cites, flag contradictions"
Known limitations
get_drug_label can silently return the wrong product
Tracked: #1 — get_drug_label silently returns wrong product on combo-drug ambiguity
The tool falls through brand → generic → substance and returns the first opening match. For ambiguous generic-name queries (e.g. empagliflozin) this can surface a combination product (Synjardy = empagliflozin + metformin) ahead of the monotherapy product (Jardiance = empagliflozin). The combo product's boxed warning belongs to the other active ingredient.
Workaround until #1 ships: query by brand name when you can. get_drug_label("Jardiance") returns the correct monotherapy label; get_drug_label("empagliflozin") does not.
This was caught the moment the server went live — the kind of failure the unit tests miss but evals against real public APIs catch. Fix is scoped for v0.2.
Tests
pytest
# 10 tests covering both tools with mocked HTTP via respx — no network calls
Tests use respx to mock NCBI + openFDA so the suite stays fast and offline.
Project layout
server/
main.py FastMCP entrypoint (stdio + http transports)
config.py pydantic-settings
tools/
pubmed.py E-utils esearch + esummary
openfda.py /drug/label.json with brand→generic→substance fallback
tests/
test_pubmed.py respx-mocked PubMed queries
test_openfda.py respx-mocked openFDA queries
agent/ LangGraph agent that consumes this server (next session)
docs/
claude-desktop-config.md Setup notes + troubleshooting
License
MIT