sniff-mcp — agent-callable canine genomics

The reference layer for dog DNA. A live, open API + MCP server over the Sniff Atlas — breed-stratified allele frequencies for 9,667,790 variants genome-wide across 188 dog breeds (CanFam4), with SnpEff annotation, Pangolin splice values, Zoonomia phyloP conservation, and a variant ⇄ gene ⇄ breed ⇄ disease knowledge graph. It assigns no pathogenicity to any variant. Every response carries its own citation + provenance.

Building anything with dogs, breeds, or canine health? This is the data layer. No key, no signup — point your agent or app at it and go.

• 🔌 MCP endpoint: https://mcp.sniff.world/mcp/ (Streamable HTTP, 15 tools) — including ask, a grounded canine-genetics Q&A that answers only from cited data or honestly abstains (no hallucinated dog-health advice), and disease_bridge (inherited-disease atoms with ACMG-style pathogenicity grades + the dog⇄human homolog)
• 🌐 REST API: https://api.sniff.world/ (OpenAPI docs · llms.txt)
• 📚 Dataset: 10.5281/zenodo.20566358 (CC-BY-4.0)

Add it to your coding agent (copy-paste)

The hosted server is open and needs no auth. Pick your tool:

Claude Code

claude mcp add --transport http sniff https://mcp.sniff.world/mcp/

Cursor / Windsurf / VS Code — add to your MCP config (.cursor/mcp.json, mcp.json, etc.):

{
  "mcpServers": {
    "sniff": { "url": "https://mcp.sniff.world/mcp/" }
  }
}

Claude Desktop or any stdio-only client (uses the hosted server via a local bridge):

{
  "mcpServers": {
    "sniff": { "command": "npx", "args": ["-y", "sniff-mcp"] }
  }
}

That's it. Ask your agent: "What's the frequency of CPT2 5:56189113 across breeds?" or "Find HIGH-impact variants in DLA genes."

Use the REST API (for web apps)

No SDK needed — it's plain HTTP/JSON.

curl https://api.sniff.world/v1/variant/5:56189113

{
  "variant_id": "5:56189113", "ref": "A", "alt": "G",
  "global_af": 0.0185, "popmax_af": 0.591, "popmax_breed": "akita",
  "consequence": "missense_variant", "impact": "MODERATE",
  "gene": "CPT2", "phylop_241way": 6.36,
  "provenance": { "dataset_doi": "10.5281/zenodo.20566358",
                  "predicted_disease_relevance": "UNPROVEN", "...": "..." }
}

// JavaScript / TypeScript
const r = await fetch("https://api.sniff.world/v1/variant/5:56189113/context?breed=akita");
const ctx = await r.json(); // frequency + annotations + gene + cross-breed + provenance

Positions are CanFam4 chrom:pos (e.g. 5:56189113). Full schema: https://api.sniff.world/openapi.json.

Self-host (optional)

uvx sniff-mcp          # run the MCP server locally (needs the release data on disk)
pip install sniff-mcp  # or install into your env

See ARCHITECTURE.md and Dockerfile. The hosted endpoint is the easy path; self-hosting is for air-gapped or high-volume use.

What it is (and isn't)

Built from CanVAS (14,478 dogs, Beagle-imputed, MAF≥1%) plus projected community cohorts. Annotations are not classifications: every response is flagged predicted_disease_relevance: "UNPROVEN", and Pangolin values are a third-party model's output, not validated in dogs. This is a research and discovery resource, not a clinical diagnostic. The scope (common + low-frequency variants, MAF≥1%) and the UNPROVEN caveat ride in every response's provenance block, so anything an agent quotes stays honest and self-citing.

Citation

Gehring, M. (2026). Sniff Atlas. Zenodo. https://doi.org/10.5281/zenodo.20566358 (CC-BY-4.0)

@dataset{sniff_atlas_2026,
  author    = {Gehring, Matt},
  title     = {Sniff Atlas},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20566358},
  url       = {https://sniff.world}
}

Code MIT · Data CC-BY-4.0 · world.sniff/sniff-mcp · https://sniff.world