Burden scoring, AlphaMissense averaging and kinship are tools rather than something an agent assembles from raw genotype dumps — which is what makes questions spanning a whole cohort answerable at all. Two deliberate choices show the same instinct: HGVSp is restricted to canonical transcripts specifically to reduce what the model has to hold, and the population tools exist so cohort structure can be established before variants are pulled. The example prompts in the README are unusually demanding, and they are the honest indication of what this is for.
A remote MCP service over the 1000 Genomes Project dataset held in the Dnaerys variant store: 3202 samples sequenced and aligned by the New York Genome Center, 2504 unrelated from the phase three panel plus 698 from 602 family trios. Variants come annotated by VEP 115 against GENCODE 49, with ClinVar clinical significance, gnomAD frequencies and AlphaMissense scores and classes.
- Variant and genotype selection — `findVariants` and `findVariantsInSamples`, filtered by coordinates, annotations, zygosity and population
- Sample-side queries: `findSamples`, `findSamplesHomozygousReference`, `findSamplesByPopulation`, `getSampleMetadata`
- Cohort computations run server-side — `computeAlphaMissenseAvg` and `computeVariantBurden`, plus `getKinshipDegree` for relatedness
- Population reference data through `listPopulations`, `listSuperpopulations`, `getPopulationStats` and `getSuperpopulationSummary`
- `getDatasetInfo` for what the store actually contains
- Returned variants carry HGVSp annotations restricted to canonical transcripts, gnomAD exome and genome frequencies, AlphaMissense scores and cohort-wide statistics; samples carry family id, gender, parental ids, relationship, children, population, superpopulation and the phase three indicator
Nothing — the hosted service is at `https://db.dnaerys.org/mcp` over Streamable HTTP, with no authentication and no user data collected. Running it yourself needs JRE 21 and a Maven build producing `onekgpd-mcp-runner.jar` as an uber-jar; the dev profile enables both stdio and HTTP transports, with HTTP on port 9000. Claude Desktop should use the stdio route against the local jar — the README notes it will not work against the local HTTP option. Apache 2.0.
Build from source — clone the repository and build it, then deploy it and point your client at the endpoint