Config

settings.toml selects which model entries in models.toml handle each task and configures local settings.

All model tasks are optional. Omitting models.chat falls back to the free hosted chat tier; embedding, rerank, image-generation, text-to-speech and speech-to-text are enabled by adding the appropriate model ids.

Example:

[models.chat]
model = "chat-default"

[preferences]
thinking = false
sounds = true
voice = false
language = "en"
persona = "default"

Fields:

models: mapping from task name (chat, embedding, image-generation, etc.) to a [[models]].id from models.toml.
preferences.thinking: show a thinking/generating indicator when the agent is composing.
preferences.sounds: play UI sounds.
preferences.voice: enable spoken replies when a TTS model is configured.
preferences.language: en or hu (affects local UI, not model behavior).
preferences.persona: default persona id to open the CLI with.

Notes:

To enable TTS/STT, add corresponding models entries in models.toml and a stt/tts block in settings.toml (see the Voice page).
settings.toml is written by the setup wizard on first run; you can edit it manually afterwards.

`kaja config fetch`

Downloads mcp.toml and models.toml from a Kaja server, backing up any existing files first:

kaja config fetch

Set services.toml’s [api] baseUrl (and secrets.toml’s [api] token, if required) before running this. On a fresh install, the fetched models.toml’s first chat-task model is also auto-filled into settings.toml’s models.chat, so a single fetch is enough to leave a fresh install bootable (a chat model you’ve deliberately chosen is never overwritten by a later fetch).

`kaja config wipe`

Backs up the whole config directory (settings.toml, models.toml, mcp.toml, services.toml, secrets.toml, tools/, personas/, datasets/) by renaming it to <dir>.bak (or .bak2, .bak3, … if backups already exist), leaving a clean slate for the next run of kaja to recreate:

kaja config wipe

Nothing is ever deleted — recover by renaming the .bak directory back if needed.

Models