Genomi: a local-first MCP runtime for querying your own genome

The problem Genomi picks: your own variant file, not population genetics

A consumer genotyping file or a whole-genome VCF is a few hundred thousand to a few million variant rows. The README puts the scale plainly: roughly 3 billion base pairs, more than 20,000 genes, and millions of observed variants per person. That is the specific problem. Not variant calling, not alignment, not cohort analysis. Genomi starts after you already have a variant file and asks what an agent can do with it locally. The stated audience is the person who owns the DNA file and wants to ask questions like medication response or inherited traits without shipping the file to a service. The framing in the README is careful about what this is not: DNA shapes risk, it is not destiny. That caveat matters because the rest of the design leans on it. Genomi is for the person who has the file, has an agent host, and wants the two connected. It is not for a bioinformatician doing cohort work, and nothing in the material suggests it does alignment or variant calling.

The Active Genome Index is the load-bearing idea

The core mechanism is a local store called the Active Genome Index, abbreviated AGI throughout the README. Genome intake is a setup step: the user hands the host agent a local VCF, gVCF, or another supported genome-source path, and the agent prepares and selects the AGI. From then on, every query reads that index. The README states that GenomiLab investigations reuse the active AGI and never ask for another genome upload or copy AGI rows into a separate database. That single sentence tells you the architecture: one parsed representation on disk, multiple consumers reading it, no second copy. Around the index sit three other pieces. Public genetics evidence is made queryable, so the agent can pull reference material rather than answer from parametric memory. Memory records what was already explored, which is what makes follow-up questions cheap. Report tools assemble the answer with citations. The privacy claim follows from this layout rather than from a separate encryption layer: the genome stays on the machine because the index is on the machine. The README also links a video titled 'Genomi knows when to say No and I don't know', which addresses the failure mode you would worry about most, an agent inventing a plausible-sounding interpretation for a variant it cannot actually resolve.

Installation runs through the agent, not through pip

The canonical path is agent-mediated. The README gives a paste-in instruction pointing at INSTALL_FOR_AGENTS.md, and states that the guide covers dependency checks, library selection, MCP registration, optional genome-source import, and verification. Python 3.10 or later is the stated floor. For hosts that already have Genomi packaged or present, the README names the update path as the command genomi install or the MCP operation genomi.install, and says the source bootstrap is only for hosts that do not have Genomi yet. That distinction is worth reading twice: if you are updating, do not re-run the bootstrap. Serving is a separate step, genomi serve, which the README describes as a long-lived MCP process. The GenomiLab section adds a constraint on that process: the same long-lived genomi serve process owns the session-scoped AGI handle, the GenomiLab domain service, the encrypted research store, and the loopback portal, and GenomiLab does not discover or launch a second embedded agent server. So the deployment shape is one server process per host, not one per feature. On a fresh Genomi home, giving the agent a genome-source path alone creates a usable local placeholder profile that the patient can rename later; only an ambiguous multi-user home needs a follow-up choice about which profile owns the genome.

What GenomiLab adds, and why the preview label is doing real work

GenomiLab is described as a patient-facing disease-investigation application built on Genomi, and it is labelled a developer preview. The design is unusual enough to spell out. The host agent owns the conversation, task lifecycle, planning, subagents, tool calls, streaming, follow-ups, resume, and cancellation. GenomiLab supplies the skill, typed patient-research capabilities, scoped authorization, durable evidence, hypothesis and brief records, and validation at those boundaries. For each new investigation the main host agent chairs a board of 2 to 5 adaptive, non-overlapping specialist subagents. The chair holds the patient conversation, authorization, private AGI reads, and canonical research commits; specialists get bounded public questions or the minimum approved evidence they need. GenomiLab records the logical board and milestone states for the portal, never raw agent messages, chain of thought, or native task identifiers. The web portal handles patient onboarding, molecular-profile updates, exact approvals, provider setup, and monitoring committed domain events, and it does not start, message, or cancel a host task. A follow-up stays in the same host task: the host records the observation, the patient approves the exact context change in the portal, and the host reruns only affected evidence before committing a superseding hypothesis and revised brief. If that reads like a lot of machinery for a preview, it is. The material does not say how the board size is chosen, how specialist overlap is detected, or what happens when a rerun invalidates a hypothesis that earlier briefs depended on.

Where Genomi is the wrong tool

The material does not claim clinical validity, and you should not read it in. A consumer genotyping array covers a small fraction of the genome, so a negative result in a genotyping file means the array did not look, not that the variant is absent. The README's own framing, that DNA is not destiny, cuts against treating any output as a diagnosis. The second limitation is structural. GenomiLab requires a current Genomi user with a query-ready AGI selected before the Research Desk opens, which means there is a real setup gate: no index, no investigation. That is a deliberate trade for privacy, but it also means a user who cannot produce a VCF or gVCF in a supported format has nothing to run. Third, the disease-investigation layer is a preview, and the README does not describe a validation corpus, a comparison against a clinical pipeline, or an error rate. A tool that is good at routing an agent to the right public evidence record is not the same as a tool that is right about your risk. Those are different claims and only the first is supported here.

The alternative is a general agent harness plus your own scripts

The obvious comparison is not another genomics product. It is the general-purpose MCP host you already have, pointed at a directory of VCF files with no domain layer. The README itself invites that comparison, asking the agent to explain why Genomi differs from other AI agent harnesses. The difference in approach is concrete: a general harness gives the model file access and lets it improvise a parsing strategy each session, while Genomi fixes a parsed index once and reuses it, adds a curated public evidence layer so answers are grounded in retrieved records rather than recall, and keeps memory of prior exploration. The GenomiLab section pushes further, arguing that its capabilities go beyond both base Genomi and a general harness because they project an approved Patient Molecular Profile, keep source-separated disease evidence, relate public findings to exact patient observations, and maintain hypotheses, counterevidence, gaps, confirmation needs, and versioned records. Whether that extra structure earns its complexity is not something the supplied material can settle. What it does settle is the shape of the choice: a general harness is cheaper to start and unbounded in what it can attempt, while Genomi constrains the task to a prepared index and an evidence layer, which is exactly what makes its answers checkable.

Licence, maintenance, and what to verify before you commit

Genomi is Apache-2.0, which permits commercial use and modification and includes a patent grant. That is a permissive licence, not a legal opinion, and it says nothing about the licences of the public genetics evidence sources the agent queries or the terms of the genome-source files you feed it. Those are separate questions and the README does not address them. On maintenance: the repository is not archived, the most recent push is dated 2026-08-31, and the only listed release is v0.1.0 from 2026-07-01. A single 0.1.0 release with a developer-preview feature layer means the interface surface, particularly the MCP operation names and the AGI layout, should be treated as movable. The README already shows one sign of that: genomi install is presented as the canonical path while the source bootstrap is retained for hosts that lack Genomi. If you build on the MCP operations, pin the version you tested and re-read INSTALL_FOR_AGENTS.md after each upgrade. The cheapest verification is to run genomi serve, point the agent at a genome-source path you control, and then ask a question whose answer you can check by hand in the raw file. If the agent cites the right record, the index and evidence layers are wired correctly. If it answers without a citation, stop there.

Editorial conclusion

Adopt Genomi if you already run an MCP-capable host, you have a VCF or gVCF you are willing to keep on your own disk, and your questions are of the form 'what does this variant in my file say'. Do not adopt it if you need clinical-grade interpretation, a second opinion on a diagnosis, or a result you can hand to a physician as a finding. Before trusting anything it produces, verify three things in your own checkout: that the Active Genome Index was built from the file you think it was, that the evidence lookups return the source records the agent cites, and that the 'No' and 'I don't know' path actually fires on a question your genome cannot answer. The GenomiLab Research Desk is a developer preview, so treat it as something to inspect rather than something to rely on.

Official sources

1. exon-research/genomi on GitHub
2. License: Apache-2.0
3. Project website
4. README
5. Releases