sniff.world

Sniff MCP — agent-callable canine genomics over the Sniff Atlas (open, CC-BY-4.0). Covers 9,667,790 common (MAF>=1%, incl ~3M at 1-5%) canine coding variants across 188 dog breeds on the CanFam4 assembly, with breed-stratified allele frequencies, calibrated ESM2 pathogenicity (AUC 0.935 vs OMIA), Pangolin splice, and Zoonomia phyloP conservation. IDENTIFIERS: positions are CanFam4 'chrom:pos' (e.g. '5:56189113' or 'CANFAM4:5:56189113' or 'chr5:56189113'). Assembly defaults to canfam4. START: for a natural-language question ('what is X', 'does breed Y get Z', 'human equivalent of W'), call `ask` — it returns a GROUNDED, CITED answer (or an honest abstain) over the fused knowledge layer (OMIA inherited diseases + dog<->human homolog bridge, AVCG variant pathogenicity grades, carrier risk, longevity, temperament, diversity). For the documented-disease atoms of a breed or disease, call `disease_bridge`. For a single variant, `ask_variant_context` returns frequency + pathogenicity + gene + cross-breed + provenance in one call. Use `variant_search` for filtered discovery, `gene_summary`/`breed_summary` for rollups, `metadata`/`breeds_in_atlas` for discovery. SAMPLE-SIZE CONFIDENCE: breed-level responses carry n_dogs, an af_ci95 (Wilson interval), and a confidence grade (high/moderate/low/very_low). The atlas is sample-skewed (median ~22 dogs/breed); an af of 0 in a 12-dog breed is NOT 'absent' (its CI may reach 0.10+). ALWAYS weight by confidence — do not report a frequency, a popmax, or a 'breed with the most/least X' claim without its n_dogs/CI. RIGOR CONTRACT: every response carries a provenance block (data DOI, evidence grade, citation). Pathogenicity outputs ALWAYS include predicted_disease_relevance='UNPROVEN' — these are computational predictions, NOT clinical diagnoses, and the resource is common-variant only (MAF>=1%). The OMIA disease layer + grounded `ask` are LIVE (cite-or-abstain; carrier != affected; educational, not diagnostic). Always surface the citation and the UNPROVEN caveat to the user.