Five interchangeable model providers, switching by voice, and the fallback chain that keeps you running.

The “Brain” is the language-model layer. Jarvis is multi-provider by mandate — no provider is hardcoded, and you choose which one answers.

The five providers

Claude (Anthropic)
Gemini (Google)
Grok (xAI)
OpenAI
OpenRouter — one key, many models

Each is configured under [brain.providers.*] and selected with [brain].primary. You only need keys for the providers you actually use.

Switch by voice

Provider switching is a first-class runtime operation, not a config-file edit:

“[wake word], switch to Gemini.”

[wake word] is a neutral documentation placeholder for the activation phrase supported by your selected wake-word provider. It is not a required product phrase.

It takes effect on your next command — no restart, no reload. The BrainManager owns the live switch.

The fallback chain

If your primary provider errors, rate-limits, or times out, the BrainManager falls through a smart chain to the next healthy provider rather than failing the command. You stay in the conversation; the swap is invisible.

The Ack tier

Separately from the deep brain, a fast Ack-Brain produces the sub-second spoken acknowledgement (see voice pipeline). It runs a small, fast model so the confirmation lands before the real answer is ready.

Workers use Claude Max

Background mission workers run through the claude-cli backend using a Claude Max OAuth session — so heavy work does not require a separate Anthropic API account. The interactive brain and the worker brain are configured independently.

Local vs cloud

Every provider class keeps a cloud-reachable default, so a keyless-local or GPU-less machine still runs entirely through the cloud. Local models are an installed-by-default upgrade whose weights download on first use — never bundled, never mandatory.