Skip to content
OdyssAI

Welcome

The Sirens promise convenience. The giants own the gates. The sea is faster if you let someone else steer. OdyssAI is for those who would rather take the helm.

OdyssAI is a stack for running frontier open-weight models on hardware you control, and for putting a real AI operating system on top of them. It is local-first: by default your prompts stay on your machines. It is not local-only: cloud models sit behind the same address, and you decide, axis by axis, when a request may cross the line.

This page tells you what the pieces are and which one you need.

Two pieces

Section titled “Two pieces”
PieceWhat it isRepoLicence
OdyssAI-XThe engine. Distributed MLX inference on Apple Silicon clusters, with an OpenAI- and Anthropic-compatible API. Three ways to put Macs to work: split one model across nodes, run one full copy per Mac and batch conversations, or serve vision models.Odyssai-eu/OdyssAI-XAGPL-3.0
CoeOSThe AI operating system. Shipped as Nemo, the desktop client, and the CoeOS box, the router that sends each request to the model proven best at that skill — on your engine, or in the cloud with your keys.Odyssai-eu/coeos · Odyssai-eu/coeos-boxMIT

Two more repositories complete the organisation:

  • Guardian — a confidential-content detection sidecar the client and the router can call before a message leaves for a cloud provider. MIT.
  • CodeOS — a sharpened, dedicated version for coding: a modified opencode tailored with our agentic workflow, where separate models review and validate the plan and the code. MIT.
Nemo · CodeOS · IDE agents · OpenAI / Anthropic SDKs
                         ↓
                  the CoeOS box          one address, one axis per request
                         ↓
                    OdyssAI-X            your Apple Silicon cluster
                         ↓
       Apple Silicon  ·  cloud providers with your keys

Each boundary is a standard protocol, so each layer is replaceable. The box does not care whether the model behind an axis is on your cluster or at a cloud provider; Nemo does not care whether it talks to a box or straight to an engine.

Which one do you need

Section titled “Which one do you need”
You wantReach forRead
A desktop assistant that routes every request to the right modelNemo + a CoeOS boxNemo
One address for all your tools — Claude Code, Aider, an SDK — with your own keys and real routingThe CoeOS boxCoeOS box
Frontier open-weight models served on your own MacsOdyssAI-XInstall the engine
All of itEngine underneath, box in front, Nemo on every deskGetting started

What is local, what is cloud

Section titled “What is local, what is cloud”

The engine is local by construction: weights on your disks, inference on your Macs, nothing phoned home. The box is local too: it runs on a machine you own, holds your keys encrypted, and sends a request outside only to the provider you bound to that axis. It ships a test suite for that promise. Nemo keeps its memory in a Markdown vault on your Mac.

Cloud providers are first-class behind the same surface. Bind the creative axis to a cloud model and legal_rgpd to a model on your cluster, and your tools never see the difference.

Licence

Section titled “Licence”

The engine is AGPL-3.0; Nemo, the CoeOS box, Guardian and CodeOS are MIT. For professions that cannot send a file to the cloud, CoeOS also exists as a paid suite on a dedicated server per client, with Guardian customised to the profession — see What CoeOS is.

Read first

Section titled “Read first”