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 language models on Apple Silicon. It is local-first — by default, your prompts never leave your machines — but not local-only. Cloud providers sit behind the same surfaces, and you decide when to cross the line.
This page is the entry point. It tells you what the stack does, how the three pieces fit, and which one you actually need.
What lives where
Section titled “What lives where”The stack has four layers. The protocols stay standard at every boundary so each layer is replaceable:
Layer 01 — Clients IDE agents · Companion UI · OpenAI / Anthropic SDKs ↓Layer 02 — Experience Companion ↓Layer 03 — Engine OdyssAI-X · Telemak ↓Layer 04 — Ground Apple Silicon · local engines · cloud providersYou can swap any layer without disturbing the others. Companion does not care whether it talks to a four-node Mac Studio cluster, a single MacBook Pro, or an OpenRouter key, as long as the engine speaks OpenAI or Anthropic.
OdyssAI-X is the cluster engine. It runs on Docker, SSH-spawns MLX runners across 1–5 Mac Studios, exposes OpenAI- and Anthropic-compatible APIs, and orchestrates models from 200 B to 700 B. If you have a multi-Mac setup with Thunderbolt 5, this is the engine.
Telemak is the mono-Mac runtime. A native Swift .app that lives in your menu bar, serves the same OpenAI and Anthropic APIs from a single Mac, and keeps several MLX models co-loaded in wired memory. If you have one machine and want 30–70 B models without Docker, Python, or SSH, this is the engine.
Companion is the chat client. React + Hono + Postgres, packaged as a Docker image. It handles conversations, projects, memory, skills, MCP integrations, and exposes itself as an MCP brain for external coding agents. The agent that lives inside Companion is called Némo — not a model, but the persona and orchestrator above whatever model you have routed.
When to reach for which engine
Section titled “When to reach for which engine”| You have | Reach for | Why |
|---|---|---|
| One Mac Studio or MacBook Pro | Telemak | .app bundle, menu-bar, in-process. ~80% of the Apple-Silicon use case. |
| 2–5 Mac Studios on TB5 mesh | OdyssAI-X | Docker orchestrator, RDMA cluster. For frontier models >200 B. |
| Both | Both | Telemak nodes enrol as providers in OdyssAI-X. Companion sees one unified catalog. |
| Just want a chat window | Companion + any engine | Companion speaks OpenAI and Anthropic. Pair it to Telemak, OdyssAI-X, Ollama, LM Studio, vLLM, or a cloud key. |
OdyssAI-X and Telemak are complementary, not competing. The typical deployment is Telemak on a workstation + OdyssAI-X on a server cluster, all fronted by Companion.
What is local, what is cloud
Section titled “What is local, what is cloud”OdyssAI is local-first. By default, telemetry stays on your LAN, model weights stay on your SSD, conversations stay in your Postgres.
It is not local-only. Cloud providers — OpenAI, Anthropic, OpenRouter — are first-class citizens behind the same OpenAI/Anthropic-compatible surfaces. The choice belongs to the operator, not the framework. Pair Companion to a cloud key on a Tuesday and back to a local cluster on a Wednesday without changing your UI, your history, or your memory.
A note on the mythology
Section titled “A note on the mythology”The names follow Homer:
- Odysseus — the wanderer. Codename for the distributed engine; still the container and repo name (
Odyssai-eu/Odysseus). - Telemak — the son who stays home. The mono-Mac runtime, ships the father’s absence forward.
- Némo — Outis / Nobody, the name Odysseus gives the Cyclops. The agent that lives in Companion.
You do not need the mythology to use the stack. It is there because naming things matters when the code outlives the changelog.
Read first
Section titled “Read first”- Install the stack → — the Configurator DMG, both deployment modes.
- Getting started → — pick your path, three commands to a first inference.
- Architecture overview → — the four layers in detail, when to reach for which engine.
- The cluster → — what distributed inference buys you, and what it costs.
- CoeOS → — the benchmark-composed “best at everything” virtual model.