dog-geroscience-mcp

mcp-name: io.github.w0lph/dog-geroscience-mcp

An MCP server that gives any agent (Claude, Cursor, other MCP clients) dog-specific aging tools that do not exist anywhere else:

The server also exposes one MCP prompt, dossier_briefing(compound, target_gene?), which instructs the client model to call intervention_dossier and write a six-section evidence briefing without adding anything the tools did not return. examples/rapamycin_briefing.md is a worked example written from examples/rapamycin_dossier.json; examples/selegiline_dossier.json shows the FOI and structured blocks populated for a marketed veterinary drug.

scripts/dossier_eval.py runs the dossier over 14 ITP-tested compounds plus two veterinary comparators and prints a coverage table (data/dossier_eval.json). It measures what the sources contain, not biological truth: for example rapamycin resolves 37 DrugAge experiments, 16 ITP rows, 152 corpus mentions and 5 candidate dog trials but no FOI summary (not a veterinary product), while selegiline resolves 1 dog lifespan experiment, 4,117 openFDA dog reports and 2 FOI summaries with 2 structured records, and carprofen 46,859 openFDA dog reports plus FOI summaries and structured records up to each search's result cap (11 and 19).

Adjacent servers this one deliberately does not duplicate: sniff-mcp (canine genomics, OMIA, breed allele frequencies), mcp-veterinary-fda (openFDA animal adverse events, Green Book), and the longevity-genie servers (Open Genes, SynergyAge, gget).

Install

No build step: the first run downloads the prebuilt database (~90 MB) from the Hugging Face Hub (w0lph/dog-geroscience-mcp-data) into a per-user cache directory, and every offline tool works from that single file.

uvx dog-geroscience-mcp                 # stdio server; downloads the database on first start
uvx dog-geroscience-mcp fetch-data      # or download it explicitly (prints the path)

Claude Desktop / Claude Code / Cursor config:

{
  "mcpServers": {
    "dog-geroscience": {
      "command": "uvx",
      "args": ["dog-geroscience-mcp"]
    }
  }
}

Docker (the root Dockerfile of the repository bakes the database into the image):

docker build -t dog-geroscience-mcp . && docker run -i --rm dog-geroscience-mcp

Environment: DOG_GERO_DATA (where the database lives; default %LOCALAPPDATA%\dog-geroscience-mcp on Windows, ~/.cache/dog-geroscience-mcp elsewhere, <checkout>/data in a source checkout), DOG_GERO_DB_URL (alternative download URL), DOG_GERO_AUTO_FETCH=0 (fail instead of downloading when the database is missing).

Build from sources

uv sync
uv run dog-geroscience-mcp build            # downloads HAGR + DAP codebooks, indexes ../corpus/data and ../foi/data
uv run dog-geroscience-mcp build --skip-download
uv run dog-geroscience-mcp build --fulltext-licences "cc by,cc0" --out dog_geroscience.sqlite   # the redistributable build
uv run dog-geroscience-mcp                  # serve from data/dog_geroscience.sqlite
uv run mcp dev src/dog_geroscience_mcp/server.py   # inspector (optional)

The build writes one SQLite file: plain tables + FTS5 indexes, the corpus's Markdown full text (all of it locally; only CC BY / CC0 articles in the published file), the full FOI records, and a meta table with every source URL and fetch time. DOG_GERO_CORPUS, DOG_GERO_FOI and DOG_GERO_FOI_STRUCTURED point the build at the inputs.

In a source checkout, Claude Desktop can run it without PyPI:

{
  "mcpServers": {
    "dog-geroscience": {
      "command": "uv",
      "args": ["--directory", "D:/OneDrive/k9/mcp", "run", "dog-geroscience-mcp"]
    }
  }
}

Design

• Ground-truth tables (AnAge, DrugAge, GenAge, DAP codebooks) are served verbatim with their source; nothing is merged or inferred.
• Network at request time is limited to the live tools (Ensembl, RePORTER, and openFDA inside the dossier); Ensembl results are cached in the database. The database itself is fetched once, atomically, and checked for the SQLite header before use.
• Query functions in queries.py are pure and tested offline; server.py is a thin wrapper.
• Licences: code MIT; HAGR data CC BY 3.0 (commercial use permitted with attribution); DAP codebooks are public GitHub files; corpus content keeps each article's own licence.

Test

uv run pytest -q
uv run ruff check src tests

Tests build a miniature database from fixtures and drive the server through the SDK's in-process client; nothing touches the network.