Overview
Avatar is the model Kaeonix runs on — trained for software engineering rather than adapted to it. How to pick a revision, override it per session, and run a different model instead.
The model
Avatar is the default model behind every Kaeonix surface: the CLI, the browser chat, and the API. It is the same model in all three, so a prompt that works in chat behaves the same way in a session, and a plan drafted in one can be handed to the other.
The current revision is avatar-1. Revisions are additive — an older id keeps resolving to the build it named, so pinning one in CI is safe.
Choosing a revision
With no argument, browse what your account can reach:
With an id, set the default outright:
The default is stored in ~/.kaeonix/settings.json, so it follows you across repositories. A project can override it for everyone by committing a model in .kaeonix/settings.json — see Configuration.
Overriding per session
To use a different model for one session without touching the default:
--model <id>Model for this session only. Bare ids resolve to Avatar revisions.avatar-1KAEONIX_MODELSame, from the environment — the form to use in CI.unsetA practical split: a strong model for the main loop, and cheaper models for subagents doing mechanical work. Subagent presets already pick sensible tiers — see Subagents.
Local models
Ollama is supported as a provider, so a model running on your own hardware needs no key and no network egress at all. This pairs naturally with the default network-denied sandbox: the agent can work on a codebase with no outbound connectivity whatsoever.