Model selection.
Veyyon works with any provider. You bring the key and pick the model. There are three model slots, and each one is set on its own — no default chain deciding for you.
Each model set on its own.
The model you talk to is separate from the ones that run in the background. Leave a background slot unset and it falls back to your interactive model.
- Interactive model — chosen live with /model or --model.
- Subagent model — one setting for the model your subagents run on.
- Compaction model — one setting for the model that compacts context.
- No default-chain — switching the interactive model never blends into the background slots.
Roles, when you want them
Optional named model assignments for specific work types (planning, review). They live in settings, scoped per profile, out of the model picker. default is not a role or a model.
Compaction, three knobs
Not a maze — a threshold, a type (handoff or snap), and a model. Long sessions keep task continuity without a wall of settings.
Any provider, or a local model.
Anthropic, OpenAI, Google, Groq, OpenRouter, Mistral, xAI, and Bedrock, plus Ollama, LM Studio, and llama.cpp found automatically. Your key talks to your provider directly.
| If you want | Try | Notes |
|---|---|---|
| Hard refactors | a strong hosted model | Better first-attempt edits and planning on deep changes. |
| Lower cost per session | cheaper subagent + compaction models | Keep your interactive model; spend less on the background work. |
| Faster iteration | a fast local or small model | Tight edit-test loops. Hashline recovery still helps weaker models. |
| To stay local | Ollama · LM Studio | Keys and weights stay on your machine. |
| CI runs | a pinned mid-tier model | Predictable cost. Pair with --approval-mode auto-edit. |
Full setup, precedence, and per-model tuning live in the handbook.