Configure Providers and models
AI Providers belong to the current desktop device. They are not part of an app and do not move with ZIP or Git.
Confirm the current access mode
Theseus shows one of two modes according to the current Qixin account entitlement:
- Qixin AI mode: only platform-provided Qixin AI models are visible; third-party configuration is hidden.
- BYOK mode: in addition to Qixin AI, you can configure OpenAI-compatible Providers or sign in to a Codex OAuth Provider.
If BYOK should be available but the settings entry is missing, sign in to the Qixin account and recheck the entitlement from AI settings. Do not edit local files to bypass the UI state.
Use Qixin AI
- Sign in with a Qixin account that has AI entitlement.
- Open System settings → AI.
- Confirm that the Qixin AI row has no authentication or network error.
- Refresh its models.
- Open AI Command Center and select an available model.
Add an OpenAI-compatible Provider
This entry is visible only in BYOK mode.
- Add a Provider under System settings → AI.
- Enter a recognizable name, service endpoint, and API key.
- Save the Provider.
- Refresh models on that row and confirm that models returned by the service are visible.
- Select a model in AI Command Center and send a read-only test question.
Model IDs come from the Provider. Do not treat example model names in documentation as a fixed availability list.
Use Codex OAuth
- Select the Codex sign-in action in BYOK mode.
- Complete the device sign-in flow shown by the UI.
- Return to settings and confirm that Codex is connected.
- Select a model under the Codex Provider in AI Command Center.
Codex OAuth is a separate Provider. It does not require an OpenAI-compatible API key or endpoint.
Choose a model
| Task | Prefer |
|---|---|
| Complex work across Pages, Graphs, and Objects | Strong reasoning and sufficient context capability |
| Reading errors and making a small repair | Fast and stable responses |
| Vision samples and image tasks | A model that explicitly supports image input |
| Explaining documentation or writing short text | A fast model is usually sufficient |
The model selector shows capabilities reported for the current Provider. A thinking level is applied only when the selected model declares support for it.
Security and troubleshooting
- Never put an API key, password, or device authorization code in chat, screenshots, documentation, or Git.
- If one Provider fails, repair that row without deleting healthy Providers.
- Save configuration before refreshing models; unsaved edits are not active configuration.
- Resolve endpoint, authentication, and network errors on the corresponding Provider row.
- Confirm both that settings can list models and that AI Command Center can select one.
Continue with Use AI Command Center.