Open-source · MIT · An AI operating system
Not the best model overall. The best model for this task. Debug goes to the debugger, prose to the writer — decided per request, from data, and shown to you every time.
What CoeOS is
A desktop client on your Mac and a router on your network. Underneath, the models are served by an OdyssAI-X engine on your own Apple Silicon, or by a cloud provider you declared with your own key. The router does not care which; it routes.
Three modes: Nemo, the everyday assistant with a memory kept in a local Markdown vault; Cowork, a document editor where the model works beside you, stage by stage, each one waiting for your go-ahead; Code, an orchestrator that plans into a task file and dispatches to a panel of agents.
Signed and notarized .dmg, Apple Silicon only. Point it at a box or an engine in
Settings → Providers; the models it can use are discovered, never typed.
One address that speaks the OpenAI and Anthropic APIs. Each request is classified on a competence
axis and relayed to the model your TMB Settings bind to it — with your keys, on your machine. Users,
tokens, quotas, a console. Claude Code, Aider, Cline, an SDK: they all call one model, CoeOS.
Nothing leaves your network except the request you chose to send to a provider. No telemetry, no kill switch. There is a test suite for that promise.
Also A sharpened, dedicated version for coding is available as CodeOS: a modified opencode tailored with our agentic workflow, where separate models review and validate the plan and the code. CodeOS on GitHub ↗
The open-source objection, up front
Trade a little quality for sovereignty and a zero bill. A good-enough model instead of the frontier one.
That's the assumption CoeOS is built to break.
Not good-enough. Better.
A single closed model is a generalist. Strong everywhere, exceptional nowhere. It drags its weak axes into every request whether the task needs them or not — and bills for all of it, including the reasoning tokens it burns when the task didn't call for deep thinking.
The CoeOS box runs the specialist for each skill — the model that scores highest on that exact axis, measured across 36 open and cloud models on 18 disciplines. Debug goes to the debugger. Legal goes to the legal analyst. The request lands once, on the right model, and returns untouched.
Result: 97.2/100 on the TMB benchmark — statistically on par with Fable 5 (96.9) and Claude Opus 4.8 (96.0) on equal-footing API runs. The point-spread is within stochastic noise; what isn't noise is the cost: 4.9× less per test than Fable 5, 1.6× less than Opus 4.8.
The routing table is data you can read. Every binding of an axis to a model comes from a score, and you can change any of it: pin a local model on the axes that must stay home, a cloud model where it wins.
Earned empirically, not claimed
Not MMLU. Not a suffix on a model card. Not a number to screenshot. Dozens of models run through the same battery, scored discipline by discipline, until a real scoreboard emerges — who actually wins where, with the receipt attached. Every routing decision traces back to a score. Not a hunch. A result.
The TMB benchmarks
42 tests across five domains — agentic execution, code, planning, reasoning, and writing. Each axis has a hand-written rubric, a single run at temperature 0, and an evaluation by Claude Opus 4.8. The result isn't "model X is good" — it's "model X scores 98 on debug and 62 on Swift." Granular enough to route each request to the model proven best at that specific skill, not the highest average.
A pilot, not an aggregator
Fuses several models per query. Fans the call out, blends the answers, bills for all of it. An algorithm guessing at consensus.
Reads the request, classifies what it is — an explicit header from the client, or a small decider model you configure — and steers it, whole, once, to the one model proven best at it. One classification. One call. The answer returns untouched, in the caller's own format, with the axis and the model named in its headers.
Routing by skill, not statistical mush. It knows what the request needs before it picks who answers.
Tailored for professionals · Legal, HR, Health
Law, HR, health: some work never leaves the office. CoeOS Pro is the paid edition of the platform — the same client and router, on a dedicated server per client, no shared tenancy — with Guardian tuned to the profession's own confidential categories on top of GDPR, and review agents that read your documents against your own doctrine. The tailoring is the product.
Drop a contract, a lease, articles of association. The agent reads it against your firm's own doctrine — your Golden Rules, versioned and auditable, not a generic prompt — and returns a conformity table: each rule marked compliant / non-compliant / to verify / not applicable, with the clause quoted, the page, a severity and a recommendation.
Near-deterministic, sources cited, refuses an unreadable document. The agent instructs; you decide — never an automated legal opinion.
Checks a company HR policy against the applicable Belgian sector agreement (CCT 200, 100…). Same conformity table, same evidence, plus HR alerts and suggested updates when a clause drifts out of bounds — a 38-hour week where the CCT caps it, a missing trial period.
One engine, many verticals. Moving from legal to HR meant changing the doctrine, not rewriting the agent. Your discipline is next.
Open-source · MIT