Introducing Genomi

Why now

Personal genomics is becoming urgent because four forces are colliding at the same time.

Sequencing keeps getting cheaper

NHGRI's sequencing-cost dataset tracks the drop from roughly $95M per genome in 2001 to roughly $562 in 2021.¹Production-scale platforms such as Illumina's NovaSeq X pushed the industry toward more than 20,000 genomes per year per instrument.² Sequencing is moving from a rare medical event toward a personal data layer.

Genomics discovery is accelerating

ClinVar now reports more than 6.8 million submissions across more than 4.5 million variants.³ The NHGRI-EBI GWAS Catalog has reported more than 625,000 curated lead associations across more than 15,000 traits.⁴ Those numbers mean a DNA answer is increasingly a moving bundle of sources, assertions, confidence levels, ancestry caveats, clinical boundaries, and sometimes disagreement.

AI is entering that research loop too. AlphaFold made more than 200 million predicted protein structures available.⁵ AlphaMissense predicted effects for 71 million possible missense variants.⁶ AlphaGenome is applying AI to regulatory variant-effect prediction.⁷ Google Research has introduced AI co-scientist systems, and the FDA approved the first CRISPR/Cas9-based therapy for sickle cell disease in December 2023.⁸⁹

Health agents need genetic context

People already ask AI about symptoms, medications, nutrition, sleep, exercise, lab results, family history, and whether some odd thing is worth worrying about. If a health agent does not understand your genes, it is missing one of the most personal layers of your biology. Genes are not destiny, but they are context.

Edge computing is becoming real

The local machine is turning into an AI runtime, not just a thin client for cloud calls. On June 1, 2026, The Guardian reported Nvidia's RTX Spark push for Windows PCs and laptops, describing chips designed to run AI agents locally rather than relying on cloud computing.¹⁰ That matters for genomics: the more useful edge AI becomes, the more reasonable it is to keep deeply personal files on your own machine and let an agent work through a local, auditable index.

The three problems

You should not have to hand your DNA to startups

DNA is not normal data. It identifies you, says things about your family, does not expire, and cannot be rotated like a password. The 23andMe bankruptcy made this less theoretical: in 2025, the FTC raised concerns about potential sale or transfer of 23andMe user data.¹¹ People should be able to use their DNA without surrendering it.

No static report can keep up

A static DNA report starts aging the moment it is generated. A variant gets reclassified. A drug-gene guideline changes. A population-frequency source adds context. A phenotype-gene relationship gets stronger, weaker, or more complicated. The genome is a living evidence problem, and we keep packaging it like a one-time document.

General AI can sound right and still be wrong

A model can explain APOE, BRCA1, CYP2C19, or lactose intolerance without knowing your answer. For a real person, the questions are more concrete:

• Do I have the relevant variant?
• Was it measured, and was the region callable?
• What genome build is this on?
• What do ClinVar, gnomAD, CPIC, PharmGKB, PGxDB, or FDA sources say?
• Is this clinical, preliminary, population-level, or not applicable?

That is not prompt work. That is tool work.

The harness

Genomi is an open source AI agent harness that turns your AI agent into a personal DNA expert. It gives your agent a local genomics workspace: a private index of your genome, evidence tools, public genetics source access, an investigation journal, and a dashboard.

The agent handles the conversation. Genomi handles the parts where guessing is not acceptable. It works through MCP-capable agent hosts, including Claude Code, Codex, OpenCode, OpenClaw, Hermes, Cursor, Cline, Goose, Roo Code, Windsurf, Claude Desktop, Gemini CLI, and other environments that can connect to an MCP server.

Active Genome Index

Your raw DNA file should not be pasted into a model. It is too large, too sensitive, and too easy to turn into a blob of context that no one can audit. Genomi parses genome files locally into an Active Genome Index: a queryable record of alleles, zygosity, quality, depth, filters, genome build, callability context, and source metadata.

Instead of asking a model to read a giant file, the agent can ask focused questions: does this sample contain rs429358, is the genotype heterozygous or homozygous, was the region covered, and is a missing finding a real absence or missing data?

Trustworthy

Evidence tools, not genetics fan fiction

Genomi currently has 89 registered operations, including the dispatcher, or 88 domain operations excluding genomi.invoke. They cover variant resolution, ClinVar matching, gnomAD frequency lookup, GWAS association review, pharmacogenomics, PharmCAT workflows, CPIC, PharmGKB, PGxDB, FDA pharmacogenomic evidence, polygenic score search, rare disease and phenotype review, HPO normalization, GenCC, ancestry context, functional genomics, sequence utilities, pathway lookup, journaling, and dashboard rendering.

That sounds like a lot because genetics is a lot. The point is not to drown the user in tools. The point is to keep the agent honest.

Roughly 30 evidence source families

Genomi connects agents to public evidence sources and local libraries across the genetics ecosystem, including ClinVar, gnomAD, HPO, GenCC, GENCODE, ENCODE SCREEN, 1000 Genomes, PharmCAT, PGS Catalog, ClinPGx, PGxDB, FDA pharmacogenomic biomarker tables, FDA pharmacogenetic association tables, GWAS Catalog, Reactome, KEGG, Human Protein Atlas, PanglaoDB, CellMarker, Open Targets, ChEMBL, PharmGKB, GeneReviews, MONDO, Orphanet, OMIM, ClinGen, and PubMed or primary literature review workflows.

Missing evidence should not quietly become negative evidence. If a source is unavailable, the answer should say that. If a claim is preliminary, it should say that. The answer should show its work or admit it cannot.

What Genomi is not

Genomi is experimental. It is for research and informational use only. It is not a diagnostic device. It does not replace a clinician, genetic counselor, pharmacist, or qualified clinical laboratory.

That caveat is part of the philosophy. DNA interpretation is hard. Many findings are uncertain. Association is not diagnosis. Polygenic scores have ancestry, calibration, and transferability limits. Pharmacogenomic evidence depends on the drug, gene, variant, diplotype, source, and clinical context. Rare-disease interpretation requires phenotype, inheritance, family history, sequencing quality, and clinical confirmation. Genomi's job is not to turn every DNA question into a confident answer. Its job is to help an AI agent show its work.

Privacy

The raw genome stays on your machine. Genomi creates the Active Genome Index locally, then lets the agent ask scoped questions against it. Public lookups use targets like rsIDs, genes, drugs, conditions, traits, and guideline questions.

Existing Active Genome Index access follows session approval rules. Project journals reject private sample evidence links in v1. Memory exports omit private evidence links unless explicitly requested and approved.

This does not make every possible workflow private. If you paste sensitive findings into a cloud model, you still need to understand that provider's policies. But Genomi changes the default shape: your DNA starts local, the agent asks for what it needs, and the evidence path is explicit.

For extra privacy, pair Genomi with a local LLM or zero-retention model settings.

Dynamic

Static DNA reports age the moment they are generated. Genomi is built for a world where the evidence keeps moving: new variant classifications, new drug-gene guidance, new source releases, new papers, and better agent skills.

With /genomi update, your agent can update the Genomi workspace it uses to sync with the latest research. When the science moves, Genomi moves with it.

With /genomi decode, the agent can assemble a self-contained local HTML dashboard across variants, ClinVar context, pharmacogenomics, ancestry reference-panel context, phenotype and inherited-risk signals, rare disease and cancer context, nutrigenomics, GWAS associations, polygenic score context, and the investigation journal.

Genomi is not a chatbot with a genomics prompt. It exposes a small base MCP surface plus a dispatcher for specialized genomics tools. Agents read capability-specific skills before calling operations, inspect result envelopes, journal meaningful findings, and route follow-up questions through the right evidence family.

The model does not need to pretend it memorized all of genetics. It needs to know how to ask the right tool, inspect the evidence, and stop when the evidence is not enough.

Docs Agent Setup

Open a local-capable AI agent and paste:

Install and configure Genomi by following the instructions here:
https://raw.githubusercontent.com/exon-research/genomi/master/INSTALL_FOR_AGENTS.md
The agent should fetch that file first, ask one question at a time, assemble a single installer command, run it, and verify the host MCP configuration.

Your genome should not be a dead file. It should be private, queryable, evidence-grounded, and alive. Genomi is our first step toward that future.