Gnomad Genetics MCP Server

What gnomad-genetics-mcp-server does

gnomad-genetics-mcp-server exposes gnomAD genetics data through MCP, with clinical significance joined from ClinVar using NCBI E-utilities. It supports population-level variant interpretation, gene constraint analysis, variant catalog searches, and coverage checks across genes, transcripts, and genomic regions.

Variant lookups return overall and ancestry-group allele counts, allele numbers, frequencies, homozygote and hemizygote counts, quality information, callset-level data, transcript consequences, in-silico predictions, and ClinVar significance. Gene constraint queries provide pLI, LOEUF with its confidence interval, observed-to-expected ratios, and Z-scores. The server also provides two data resources for individual variant and gene-constraint records, plus a guided gnomad_variant_triage prompt that orders population-frequency, gene-constraint, and callability checks.

How it works

The server can run over STDIO or Streamable HTTP. A public hosted endpoint is available at https://gnomad-genetics.caseyjhand.com/mcp, and the package can be launched locally through npm tooling. MCP clients invoke dedicated tools for single-record lookups, filtered lists, coverage analysis, and ClinVar searches.

List operations return analytical row sets rather than stable resource documents. Those results can be staged in a canvas and inspected with gnomad_dataframe_describe; gnomad_dataframe_query then runs a read-only SQL SELECT across the staged tables. An optional drop operation removes a named table to recover canvas memory.

Setup and configuration

The repository provides installation links for Claude Desktop, Cursor, and VS Code. A local STDIO setup can use the npm package through npx, including the package's latest tag, while the hosted Streamable HTTP endpoint can be configured as the MCP server URL. The project also publishes a container image through GitHub Container Registry.

Configuration includes GNOMAD_MAX_VARIANT_BATCH, which raises the per-call variant limit, and GNOMAD_NORMALIZE_IDS, which enables case-insensitive rsID matching and conversion of colon-separated coordinate identifiers. GNOMAD_DATAFRAME_DROP_ENABLED=true enables the canvas table-drop tool; it is disabled by default.

Tools and capabilities

• gnomad_get_variant accepts variant IDs or rsIDs and returns detailed population, callset, consequence, predictor, QC, and ClinVar information. It supports batches of up to 25 IDs by default.
• gnomad_get_gene_constraint looks up loss-of-function constraint by HGNC symbol or Ensembl gene ID.
• gnomad_list_gene_variants lists variants for a gene, transcript, or region, with consequence, frequency, and QC filters and offset pagination.
• gnomad_get_coverage reports mean and median depth plus the fraction of samples above depth thresholds for each callset track.
• gnomad_search_clinvar returns gene-level classifications, review status, conditions, submission counts, and compatible gnomAD variant IDs.
• gnomad_dataframe_describe and gnomad_dataframe_query support schema inspection and read-only SQL over staged results. gnomad_dataframe_drop is available only when explicitly enabled.

Limitations and notes

Upstream-backed tools and resource reads share a fixed 45-second deadline covering queue waits, retries, response bodies, and chained requests. A variant batch shares that deadline across its items, so completed results can remain available while unfinished primary items report an upstream timeout.

Coordinate variant IDs do not carry a genome build. The server can report a possible match on the other build, but it does not rewrite or retry the identifier automatically. Mitochondrial variant IDs are unsupported. ClinVar failure can leave a valid population result available with a null ClinVar value and an availability flag.

The dataframe SQL interface is read-only and depends on tables staged by list tools. Its table-drop operation is opt-in. Dataset behavior also varies: gnomAD R4 provides joint exome-and-genome population views, while older datasets expose carried callset data without joint populations.