Companion — getting started

The house the son comes home to. Open the door, pair an engine, start talking.

Companion is the chat client. It ships no model of its own — you pair it to an engine (a local cluster, a single Mac, or a cloud key) and it gives you a consistent window with history, memory, projects, skills and MCP on top.

1. Get Companion running

Companion is installed for you by the OdyssAI Configurator (profile Serveur) — see Install the stack. After that it’s serving on http://<server-ip>:3100.

Only have one Mac and Telemak? You still need Companion for a chat window — run the Configurator’s Serveur profile on that same Mac (or any Mac on the LAN); it brings up Companion without requiring a cluster.

2. Create your account

Open http://<server-ip>:3100. On a fresh install the first screen is Create your operator account — pick a name, email and password. That account is the admin; change the default password on first login.

3. Pair an engine

Settings → Infrastructure → Engine.

You have	Add this endpoint
An OdyssAI-X cluster	`http://<server-ip>:8000`
A Telemak Mac	`http://<telemak-ip>:8003`
A cloud key	Add Cloud Provider → paste OpenRouter / Anthropic / OpenAI key
Ollama / LM Studio / vLLM	their local OpenAI endpoint

Click Test endpoint. Companion reads the engine’s capability contract (/.well-known/inference-engine.json), learns which models support tools, vision and thinking, and loads the catalog. You can pair several engines at once — Companion keeps a separate catalog per engine and lets you route per message.

4. Send your first message

Open a new chat, pick a model from the model picker (top of the conversation), and type. You get the full conversation surface: edit / regenerate, attachments, voice, code-block helpers.

The footer under each answer (the StatsRow) shows TTFT, tokens/s, token counts, cache hits, and the model that actually answered — handy when you route through CoeOS, which shows e.g. CoeOS · python — MiniMax-M3.

5. Turn on memory (optional)

Companion remembers across conversations through Némo, a per-user knowledge graph. Toggle it in Settings → Memory, or per-project. The agent can append to it on its own (companion_remember), and you can curate it by hand. Per-project memory keeps a project’s facts scoped to that project. See Memory.

6. Where to go next

Organise chats into Projects that share a system prompt and memory.
Extend the agent with Skills and MCP servers (Notion, Linear, GitHub, Tavily, your own).
Turn Companion into a brain for an external coding agent with Agents tokens — Cline, Continue.dev, Claude Desktop and Cowork can call back into Companion’s memory and tools.

A note on Némo

The assistant inside Companion calls itself Némo. Némo is not a model — it’s the persona and orchestrator above whatever model you’ve routed: the memory, the relationship, the editor’s voice. The model is the substrate; Némo is what the substrate says when you address it.