Models & providers
Whose AI the agent runs on — providers your organization supplies, keys of your own, and the defaults a machine reaches for.
Your agent needs at least one AI provider to think with. Hoshi never resells inference: providers connect with the credentials of your organization or your own, and inference is billed by the provider directly.
Organization providers and your own
Providers come from two places, both visible in Customize → AI providers.
- Provisioned by your organization — the org connects a provider once, and every machine in it gets the models.
- Connected by you — your own subscription or API key. A credential you connect yourself takes priority over the one from the organization for that same provider.
To connect one, pick the provider and follow its method: most take an API key, some offer a sign-in flow of their own. The key is stored on your machine and never leaves it. Disconnecting your own credential reverts the provider to the key from the organization, where there is one; disconnecting a provider entirely makes its models stop working in chat.
Choosing which models show up
Each provider exposes many models, and you rarely want all of them in the picker. Manage models hides or shows them individually — the composer offers only what you leave visible.
Defaults, per machine
Customize → Preferences sets what the machine reaches for when you do not choose explicitly.
- Default model — every new session starts on it.
- Lightweight model — small utility work, such as naming a session.
- Heavy-duty model — the hardest delegated work; falls back to the default when unset.
- Default agent — the agent new sessions start on.
You can still switch model and agent per session from the composer at any time.