Ncbi Variation MCP server

@pipeworx/ncbi-variation

dbSNP refSNP records and HGVS ↔ SPDI ↔ rsID normalization from NCBI Variation Services — the service that turns any spelling of a human genetic variant into the canonical coordinates the rest of genomics keys on.

Part of Pipeworx — an MCP gateway connecting AI agents to 1679+ live data sources.

Tools

record: genomic placement on the build you ask for, HGVS genomic/transcript/ protein forms, SPDI, gene context, ClinVar clinical significance, and population allele frequencies (gnomAD, ExAC, 1000 Genomes, TOPMED, ALFA). Answers "what is rs113488022".

• variationrefsnp(rsid, assembly?, maxfrequencies?) — the full refSNP

SPDI alleles, and validate that the HGVS is well-formed at all.

• variationhgvsto_spdi(hgvs) — normalize an HGVS expression to contextual

VCF or pipeline → the rsIDs other databases key on.

• variationspdito_rsids(spdi) — the reverse: coordinate + alleles from a

Auth

Keyless. NCBI asks unauthenticated clients to stay under ~3 requests/second; the pack sends an identifying User-Agent.

Data sources

• — refSNP record.
• — HGVS → SPDI.
• — SPDI → rsIDs.

Traps

GRCh37 vs GRCh38 is the whole game. One refSNP carries placements on both builds at different coordinates — rs113488022 (BRAF V600E) is 7:140753336 on GRCh38 and 7:140453136 on GRCh37, 300kb apart. A caller who takes a coordinate from one build into a dataset annotated on the other gets "not found", which reads as "this variant does not exist". So assembly is an explicit argument (default GRCh38), every response states assemblyrequested and assembliesavailable, and an assembly with no placement says so in assembly_note instead of coming back as an empty array.

SPDI is 0-based, HGVS is 1-based. NC000007.14:140753335:A:T and NC000007.14:g.140753336A>T are the same variant. The off-by-one is the most common reason variationspdito_rsids returns nothing.

A retired rsID is not an error. dbSNP merges rsIDs; a merged one returns mergedsnapshotdata with no primarysnapshotdata at all. The pack detects that and returns status: "merged" plus the rsID to re-query (rs3735962 → rs328), rather than an empty record.

Reference alleles are in the allele list. dbSNP lists the reference as an allele whose deletedsequence equals its insertedsequence. Filtering those out is what separates alternate_alleles from noise.

The raw record is huge. rs328 carries thousands of citations and dozens of placements. The pack summarizes and caps (max_frequencies, 25 citations) — counts are always reported alongside so a truncation is visible.

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "ncbi-variation": {
      "url": "https://gateway.pipeworx.io/ncbi-variation/mcp"
    }
  }
}

What this endpoint actually serves

tools/list at https://gateway.pipeworx.io/ncbi-variation/mcp returns the tools in the table above plus the shared Pipeworx meta-tools — askpipeworx, discovertools, search_within, remember/recall and the rest of the gateway-wide set. So the tool count you see is larger than this table: a single-pack endpoint currently lists roughly 30 shared tools alongside the pack's own. The connection's initialize response states its exact scope, and is the authoritative answer for a given day.

Shortened. The full README is on GitHub.