Engines and mixing
One interface for OpenRouter, Anthropic, Gemini, local servers and CLI agents — chosen per stage.
In short
An engine is where a model call goes: a cloud API, a local server, or a command-line agent. Thesivor connects to several, and you can use a different one for each agent and stage — a cheap model to interview you, a strong one to draft, a local model to read papers.
1Add an engine in Settings
Settings → AI engines lists the providers. Add a key for a cloud engine, point at a local server, or install a CLI agent and let it sign in itself.
2Pick models per tier
Settings → Models sets a cheap and a strong default. Per-agent overrides live in the boardroom sidebar, the vault's Agents & models control, the notebook, and the editor's AI panel.
3Mix freely
Every stored choice is engine::model. A bare model id means OpenRouter, so choices saved before engines existed keep their meaning.
For technical users
Every engine fails with the same GatewayError, so probes, refused-model marking, visible retries, switch-and-retry, step progress and Stop behave identically. Switch-and-retry suggests a tested model on the same engine first, then another configured engine. Credit is shown only for OpenRouter, the one engine that reports it. The header dot shows the worst state among the engines in use. CLI flags come from the installed tool's own --help, never from memory.