Genome Computer for Codex

Today we’re launching Genome Computer for Codex.

It’s a remote MCP server that lets Codex, OpenAI’s coding agent, work with your genome. Connect it once, sign in with your Genome Computer account, and Codex can inspect, query, validate, and build with your authorised genomic data.

Anyone with a Genome Computer account can connect it today.

Why Codex

Most consumer genomics has stopped at the answer. You log into a portal, you read a PDF, the tab closes, and the next time you want to know something about your biology you start from zero.

Codex changes the surface. In ChatGPT you can ask a question about your genome and get an answer. In Codex you can ask for a tool that answers that question today, keeps the logic visible, and can be adapted tomorrow when the question changes.

Chat turns your genome into an answer. Codex turns it into personalisation you keep.

This matters because a genome is not a one-time question. It’s a context you carry for life, and the things you’ll want to know about it change as your circumstances change and as the science moves. The right interface is one that lets you generate personalised tools and agents against your own biology, on demand — software that’s grounded in your data, not a population average, and that keeps working long after the conversation ends.

What we’re excited to see

The Codex surface is general. People will build things we haven’t thought of, for themselves, against their own genomes.

The shape we’re most excited about is agents that watch a genome over time.

A literature watcher polls PubMed, bioRxiv, and medRxiv, joins each new paper against your bundle, and surfaces only the rare paper that touches a variant you carry — a new study proposes rs1234567 in MTHFR may modulate methotrexate response; you carry this variant. Most weeks, nothing. That is the point.

A trial matcher watches ClinicalTrials.gov against your actionable variants and phenotypes. Especially valuable for rare-disease carriers and families, who today rely on advocacy networks and luck to know when a relevant trial opens.

A variant watcherre-runs interpretation when ClinVar or CPIC ships a release and surfaces meaningful changes — your BRCA2 variant moved from VUS to Likely Benign in ClinVar 2026.06. Today, when classifications change, most people are never told.

These are tractable because of two things. The bundle is structured and versioned, so an agent can say exactly what changed and why. And access is scoped per agent, so you can grant a literature watcher read access to your variant table without giving it anything else, and revoke it with one click.

Why .genome

.genome/1.0 is the format that makes the surface honest.

Every interpretation is in the bundle. Every reference — ClinVar, CPIC, PharmGKB, ACMG — is pinned to a dated version. Every pharmacogenomic phenotype is computed against guideline versions, not inferred at query time. The agent doesn’t have to guess the schema or hallucinate a rule. The schema is there. The provenance is there. The rules are there.

An LLM uses 3–7× fewer tokens reasoning about a .genome bundle than the equivalent annotated VCF, and produces 10–20× better interpretive accuracy on standard genomic-reasoning evaluations.

The bundle hosted under your account also keeps getting newer. When CPIC publishes new dosing guidance, when ClinVar reclassifies a variant, when ACMG adds an actionable gene, we re-run interpretation against the latest references and publish a new dated bundle. Previous versions stay archived. Anything connected to the MCP reads the current one.

A downloaded .genome is a snapshot. A hosted .genome is a snapshot that keeps getting newer.

The full case for the format is in introducing .genome.

From saliva or an existing file

You don’t need a genome to start.

We sequence from saliva in our own wet lab — CLIA, CAP, and NATA accredited, ISO 15189 certified. The result is a .genome bundle hosted in your account.

If you already have a 23andMe, Ancestry, or whole-genome sequencing file, we convert it. The output is the same .genome bundle, with the same structure Codex can work against.

Either path lands you in the same place: a bundle that’s queryable, current, and yours.

Privacy

Genomic data is not ordinary user data.

Access is account-authorised, scoped, and revocable per agent. The default surface exposes structured context and tool results, not a blanket dump of raw genetic files to every connected client. You remain the permissioning layer.

Genome Computer operates under Humankind, a Delaware Public Benefit Corporation. The full argument is in the Genome API post.

Get started

Connect Genome Computer to Codex at genome.computer.

Your genome is not a report. It is a substrate for personal software.