Set up AI engines
Choose a provider, add a key or sign in a CLI, and mix engines per stage.
In short
Thesivor does not include a model. You connect one — a cloud API with a key, a local server with no key, or a command-line agent that signs itself in. You can use a different engine for every stage and every agent, and nothing needs one to open a project or compile LaTeX.
Which engine should I use?
| If you want… | Use |
|---|---|
| The quickest free start | OpenRouter with a free model (daily limits apply) |
| Everything on your own machine | A local server (Ollama or LM Studio) |
| Strong long-form writing | Anthropic API or Gemini |
| To use a ChatGPT plan you already pay for | Connect ChatGPT through the Codex CLI |
| To use an existing CLI subscription | Claude Code, opencode or Antigravity (agy) |
Engines
- OpenRouter — one key, free models available, the app can auto-pick one.
- Anthropic API — your own key; strong writing.
- Google Gemini — native API; long context and images.
- A local server — Ollama, LM Studio, or any OpenAI-compatible endpoint.
- Claude Code — the CLI agent; uses its own sign-in.
- Codex / Connect ChatGPT — sign in to ChatGPT through the Codex CLI.
- opencode — the CLI agent; uses its own sign-in.
- Antigravity (agy) — Google's agent CLI; uses your Google sign-in.
For technical users
One interface (Engine) and one resolver back every provider; a stored choice is engine::model (a bare id means OpenRouter, so older values keep their meaning). Every engine fails with the same error type, so probes, refused-model marking, visible retries, switch-and-retry, step progress and Stop work identically. The header dot shows the worst state among the engines in use. Keys live in the OS secure storage; CLI agents use their own sign-in and Thesivor never reads their credentials.