Skip to content

Engine pairing

Companion ships no model. You decide what’s behind the chat window — and you can change your mind any Tuesday.

Pairing is how Companion learns where to send inference. You can pair several engines at once; each keeps its own catalog entry and you route per message.

Settings → Infrastructure → Engine → Add.

You haveEndpoint to add
OdyssAI-X clusterhttp://<server>:8000
Telemak Machttp://<telemak-ip>:8003
Ollama / LM Studio / vLLMtheir local OpenAI endpoint
A cloud keyAdd Cloud Provider → OpenRouter / Anthropic / OpenAI

Click Test endpoint. Companion fetches the capability contract, learns the per-model capabilities, and merges the models into the unified catalog.

How Companion actually reaches a model. The default is the first one; the others exist for compatibility.

RailWhat it doesWhen
Gateway (default)Talks directly to OdyssAI-X / Telemak over their native OpenAI + Anthropic surface. Capabilities from the contract, inference from the engine.The normal path. Local clusters and Telemak.
HybridCapabilities from the engine contract, inference proxied through LiteLLM.Rare — only when a third-party rail needs it.
LegacyEverything through LiteLLM.Compatibility fallback only.

The gateway is the rail. OdyssAI-X is the inference path; LiteLLM is a legacy fallback, not the centre. If a screen or doc frames LiteLLM as central, it’s stale — pair to the engine directly and stay on the gateway.

Pair a cluster and a Telemak and a cloud key, and the model picker shows all of them with the engine each model belongs to. Route per message: a cheap local model for drafts, a frontier cloud model for a hard problem, a Telemak coder for code — same conversation, same history, same memory.

Pairing changes where inference runs, nothing else. Your conversations, memory and projects live in Companion’s Postgres regardless of which engine answers. Switch from a local cluster to a cloud key and back without migrating a single byte of history.