DeepDNA MCP server

A remote Model Context Protocol server that gives AI agents cited pharmacogenomic and gene-variant lookups, so the model quotes CPIC and DPWG guidance instead of recalling it from memory. This repository also holds a local package that reads a raw DNA file on the user's own computer; nothing from the file is sent to DeepDNA.

https://deepdna.ai/mcp

Streamable HTTP, stateless, no API key, no account. Listed in the official MCP registry as ai.deepdna/knowledge.

Educational information only, not medical advice. The server reports what curated sources and guidelines say about a gene, a variant or a drug–gene pair. It does not diagnose, prescribe or recommend a dose.

Tools

Two prompts, medication_review and explain_variant, package the common questions for clients that show MCP prompts.

Coverage today is 12 genes: CYP2C19, CYP2C9, CYP2D6, SLCO1B1, VKORC1, UGT1A1, G6PD, MTHFR, COMT, HFE, APOE and BRCA1, with 51 curated drug–gene pairs across 11 of them: 40 cite a CPIC or DPWG guideline, the rest a drug label or the primary literature; the basis field in check_medications says which (guideline is null, or "drug label", when no guideline exists). It is a small, reviewed set, not a mirror of PharmGKB/ClinPGx. Every record lists its sources. The level field (the guideline's evidence grade) is null for every pair today.

Connect

Claude Code

claude mcp add --transport http deepdna https://deepdna.ai/mcp

Cursor (~/.cursor/mcp.json) and other clients that use an mcpServers file:

{
  "mcpServers": {
    "deepdna": { "type": "http", "url": "https://deepdna.ai/mcp" }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "deepdna": { "type": "http", "url": "https://deepdna.ai/mcp" }
  }
}

Claude Desktop and claude.ai: add https://deepdna.ai/mcp as a custom connector.

Try it without a client

curl -s https://deepdna.ai/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"lookup_drug_gene","arguments":{"gene":"CYP2C19","drug":"clopidogrel"}}}'

The answer carries the CPIC recommendation for clopidogrel by CYP2C19 metabolizer group, the sources it was written from, and the disclaimer. The same data is available as plain REST, documented in the API docs and the OpenAPI spec:

GET https://deepdna.ai/api/v1/pgx/CYP2C19/clopidogrel
GET https://deepdna.ai/api/v1/variants/rs4244285

Read a raw DNA file on your own computer

The remote server takes no genotype: its tools have no field for one. For a raw DNA file there is a local package in this repository. It reads the file on the user's computer and reports what it shows at each position DeepDNA curates, with the cited records. Nothing from the file is sent to deepdna.ai: the package downloads the same public records for every user, and lookup_drug_gene sends only the gene and drug names. What read_dna_file returns (the letters and counts at the curated positions) becomes part of the conversation with the AI model your client uses, under that provider's terms.

Claude Code (needs uv; the package is on PyPI):

claude mcp add deepdna-local -- uvx deepdna-mcp@0.1.1

Clients that use an mcpServers file (if the client cannot find uvx, put its full path, from which uvx, in command):

{
  "mcpServers": {
    "deepdna-local": { "command": "uvx", "args": ["deepdna-mcp@0.1.1"] }
  }
}

• Each position gets one of four states: read, no_call, not_in_file, or unexpected_letters (letters that fit neither the record nor its complement). A missing or unreadable position is unknown, never normal.
• For positions the record marks as countable (no phasing or other variants are needed to read them), it gives how many copies of the variant allele the file shows, with the record's text and sources.
• genes limits the answer to what the user asked about. APOE, BRCA1 and HFE are disease-risk genes: they are read only when named in genes, so the user chooses what to learn.
• sex is needed only for some chromosome X positions (G6PD).
• It checks the file's strand first: at least three positions that are not A/T or C/G must agree, and none disagree. A/T and C/G positions, whose two letters look the same on either strand, are counted only on a confirmed strand.

A count at one position is not a metabolizer status, a diplotype, a diagnosis or a dose. The tools tell the model never to suggest starting, stopping or changing a medication: that stays with the prescriber. VCF files are not supported.

Try it with: "My 23andMe file is at /Users/me/Downloads/genome.txt. What does it show at the CYP2C19 positions DeepDNA curates, and what do the sources say about clopidogrel?" (on Windows, a path such as C:\Users\me\Downloads\genome.txt).

What it is not

• The remote server does not read DNA files. Parsing on DeepDNA's servers is specified in the API but is in private beta and not deployed. A remote lookup carries a gene symbol, an rsID or a drug name. Its tools have no genotype field, and an undeclared argument is refused and not used. An agent that reads a raw DNA file itself compares the user's genotype with the reading block, outside DeepDNA. The reading block gives the reference and variant alleles on the plus strand of GRCh37, as 23andMe raw data reports it.
• Neither server calls phenotypes or doses. Turning a genotype into a metabolizer status needs phasing and the other variants in the gene, which a single lookup or a chip file does not give. Tools such as PharmCAT do that from a VCF.
• Remote usage is counted, people are not. Each request to the remote server is counted (method, tool, the lookup label described below, whether there was an answer, the client's name and version, its user-agent, a country code). IP addresses, request bodies and medication lists are not stored. A value typed into a declared field is recorded only as a short lookup label: an rsID in the form rs123, a gene from a fixed list of known pharmacogenes, or a drug DeepDNA covers; anything else, whatever it is, is recorded as "(other)". Please don't type a personal genetic result into these fields. The one exception is request_beta_access: it keeps the email and use case the user agrees to send, the signup metadata listed in the privacy policy (time, source, country code and the client's name), and a hashed IP for one hour to limit abuse. The local package's downloads are ordinary requests to the public API, with the user-agent deepdna-mcp-local/<version>.
• A consumer chip is not a clinical test. Arrays are reliable for common variants and much less so for rare ones; anything that could affect a health decision needs a validated test and a clinician.

Links

• Docs: https://deepdna.ai/docs/#mcp
• Walkthrough with real outputs: https://deepdna.ai/blog/pharmacogenomics-mcp-server/
• Worked example in Python, where your own code goes from a raw DNA file to CPIC phenotypes (neither server does this step): https://deepdna.ai/blog/pharmacogenomics-raw-dna-python/
• Terms: https://deepdna.ai/terms/ · Privacy: https://deepdna.ai/privacy/

This repository holds the documentation and the local package; the remote server is hosted by DeepDNA. Questions and bug reports go in the issues; please don't post a raw DNA file, genotypes or medications there.

_{Written with AI assistance.}