documentation · models
Models, routing and BYOK
The model is a component, not the product. Skyber lets you pick one per conversation, tune how it behaves, or point the workspace at an endpoint you control.
Per-conversation model
Each conversation stores its own model choice, temperature and maximum response length. Change them mid-thread and the next turn uses the new settings — memory, persona and history carry over unchanged.
Modes
Deep Think
Raises the reasoning budget for hard problems and slows the reflexes. Best for architecture, debugging and analysis.
Research
Adds live web search to the turn, so answers cite what is currently on the internet rather than training data alone.
Vision
Routes images and scanned documents through a multimodal path for OCR, handwriting and diagram reading.
Advanced
Unlocks the heavier workflow tooling and the alternate workspace theme for long working sessions.
Bring your own endpoint
The provider settings page accepts any OpenAI-compatible endpoint — a hosted provider, a gateway, a private inference server or a local model exposed over HTTP. Supply the base URL, the API key and the model identifier, and Skyber checks the connection before you rely on it.
- Keys are stored server-side and used only to make requests on your behalf.
- Prompt traffic goes to the provider you configured; you own that relationship and its terms.
- If a selected model cannot handle an attachment, the request falls back to a multimodal path rather than failing silently.