What CoeOS is

One client on your Mac. One router on your network. Every request lands on the model proven best at that task — local when it has to stay local, cloud when that is the better tool — and the answer comes back with the name of the model that wrote it.

CoeOS is two pieces, shipped as two open-source repositories:

Piece	What it is	Repo	Licence
Nemo	The desktop client for macOS: chat, cowork on documents, a coding mode with a panel of agents. Signed and notarized `.dmg`.	`Odyssai-eu/coeos`	MIT
The CoeOS box	The router: one address that speaks the OpenAI and Anthropic APIs, classifies each request on a competence axis, and relays it to the model your TMB Settings designate — with your keys, on your machine.	`Odyssai-eu/coeos-box`	MIT

Underneath, the models are served either by an OdyssAI-X engine on your own Apple Silicon Macs, or by a cloud provider you declare with your own key. The box does not care which; it routes.

The idea

A single frontier model is a generalist: strong everywhere, exceptional nowhere, and it drags its weak axes into every request. CoeOS runs the specialist for each skill instead — the model that scores highest on that exact axis in the TMB Scoreboard. Debugging goes to the debugger, a GDPR question to the legal analyst, a plan to the planner. One classification, one call, the answer returned untouched.

The ranking behind the routing is not a public leaderboard. It is the TMB benchmark: real tasks, hand-written rubrics, one run per model, scored axis by axis. The routing table is data you can read and edit, not a setting you have to trust.

How a request flows

Nemo, or any OpenAI / Anthropic client
        │  "model": "CoeOS"
        ▼
the CoeOS box
        │  1. explicit x-coeos-axis header, or the decider model you configured,
        │     or the default axis
        │  2. axis → the model bound to it in your TMB Settings
        ▼
an OdyssAI-X engine on your LAN     or     a cloud provider with your key
        │
        ▼
the answer, with x-coeos-axis / x-coeos-model / x-coeos-provider headers

Three things follow from this shape:

Nothing leaves your network except the request you chose to send to a cloud provider. No telemetry, no kill switch, no mandatory call home. The box ships a test suite for that promise.
Routing is visible. Every response names the axis and the model that served it. Nemo shows it in the model picker; a curl shows it in the headers.
Local and cloud are the same surface. Bind an axis to a model on your engine today and to a cloud model tomorrow; your tools do not change.

What you install

Just the client. Install Nemo, point it at a CoeOS box on your network or at an OdyssAI-X engine, and use it.
Client and router. Run the CoeOS box with docker compose up -d, give it your OpenRouter key or the address of your engine, and every tool you own — Nemo, Claude Code, Aider, an SDK — talks to one address.
The whole stack. Add an OdyssAI-X cluster underneath for frontier open-weight models served locally.

Guardian

Guardian is the confidential-content detector of the stack: a small sidecar that reports personal data, secrets and business-sensitive content in a message before it leaves for a cloud provider. It detects; the policy — warn, keep local, block — belongs to the caller.

The CoeOS suite for regulated professions

For professions that cannot send a file to the cloud — legal, health, HR and the like — CoeOS is also available as a paid suite: the same client and router, operated on a dedicated server per client, with Guardian customised to the profession’s own confidential categories on top of GDPR. No shared tenancy. Pricing is not public yet; write to hello@odyssai.eu.

Also: CodeOS

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, each routed by the CoeOS box.