pubmed-mcp

An MCP server for medical and biomedical literature research — 17 tools covering PubMed, Europe PMC, full-text retrieval, MeSH, citations, clinical trials, Semantic Scholar, OpenFDA and DrugBank.

One file, deps: FastMCP + httpx only. Works out of the box — all API keys are optional (they raise rate limits and enable a couple of extras). Read-only: it never writes to upstream services.

17 tools · optional API keys · read-only · works with any MCP client (Claude Desktop, Claude Code, Cursor, Windsurf, Hermes, ...).

The server itself is a single Python file: server.py — everything else in this repo is documentation, packaging and tests.

What it gives your assistant

Related MCP server: pubmed-evidence

Quick start

No installation needed — run it straight from GitHub with uv:

uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp

or from a clone:

pip install -r requirements.txt
python server.py

Client setup

Claude Desktop — merge into claude_desktop_config.json:

{
  "mcpServers": {
    "pubmed": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/BerkayCelk/pubmed-mcp", "pubmed-mcp"]
    }
  }
}

Claude Code:

claude mcp add pubmed -- uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp

Cursor (~/.cursor/mcp.json) and Windsurf: same JSON shape as Claude Desktop. Hermes: add to your config.yaml under mcp_servers (see examples/hermes_config.yaml).

Ready-made config files live in examples/.

For AI agents (automated setup)

Written so an AI coding agent (Claude Code, Codex, Cursor, Hermes, ...) can install and verify this server on its own.

1. Install command (no clone needed):

uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp

2. Register with a client:

• Claude Code:

  claude mcp add pubmed -- uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp

• Claude Desktop — merge into claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows, ~/.config/Claude/ on Linux):

  {"mcpServers": {"pubmed": {"command": "uvx", "args": ["--from", "git+https://github.com/BerkayCelk/pubmed-mcp", "pubmed-mcp"]}}}

• Cursor — same JSON in ~/.cursor/mcp.json (project scope: .cursor/mcp.json).

• Hermes — see examples/hermes_config.yaml.

• Any other MCP client — register this stdio command: uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp

3. Verify (no client needed):

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' | uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp | head -c 400

Expect JSON containing "serverInfo":{"name":"PubMed".

4. Usage notes for agents:

• Literature questions -> pubmed_search, then pubmed_fetch / pubmed_summary for the hits; use pubmed_europepmc_search for preprints and grey literature.

• Full text -> pubmed_fetch_fulltext (chain: PMC → Europe PMC → Unpaywall; set UNPAYWALL_EMAIL to enable the last step).

• Drug safety -> openfda_* tools (no key needed); drugbank_search requires a free DRUGBANK_API_KEY.

• Batch PMIDs whenever possible (up to 50 per call) and don't remove the built-in throttles — these are shared public APIs.

5. Copy-paste prompt for users:

Set up the PubMed MCP server for me: register uvx --from git+https://github.com/BerkayCelk/pubmed-mcp pubmed-mcp as a stdio MCP server named pubmed in my client config, restart/reload the client, then verify by calling the pubmed_search tool with a small test query.

Optional API keys

Everything works without keys. Keys only raise rate limits or enable extras:

Full text chain & rate limits

pubmed_fetch_fulltext tries, in order: NCBI PMC → Europe PMC → Unpaywall, returning structured sections (Introduction, Methods, Results, ...) when available.

The server throttles itself: 0.35 s between NCBI calls (0.1 s with a key) and 3 s for Semantic Scholar. This wraps shared public APIs — please keep the throttles. Data sources: NCBI, Europe PMC, OpenFDA, ClinicalTrials.gov, Semantic Scholar, Unpaywall, DrugBank.

Development

• Run from a clone: uv run --with fastmcp --with httpx python server.py

• Smoke test (spawns the server over stdio, a few live calls — keep it gentle): uv run --with fastmcp --with httpx python smoke_test.py

• Packaging check: uv build

Disclaimer

Unofficial community project — not affiliated with or endorsed by NCBI/NLM, Europe PMC, the FDA, ClinicalTrials.gov, Semantic Scholar, OurResearch, or DrugBank. For research assistance only; not medical advice.

License

MIT — see LICENSE.