Model picker

/model opens a picker backed by a live, dynamic catalog — not a hardcoded list shipped at release time.

How the catalog is built

The CLI layers three sources, in order, and takes the richest one that's available:

1. A baked floor bundled with the package — always works, even fully offline.
2. A cached fetch of models.dev at ~/.visvoai/cache/models.json — refreshed opportunistically.
3. If neither the network fetch nor the cache succeed, the CLI silently falls back to the baked floor rather than blocking startup.

That stack resolves to thousands of models across 100+ providers — the figure moves every time models.dev adds an entry, which is the point of not shipping a hardcoded list. The picker doesn't show you all of them, though — see Providers & model compatibility for the full table of named providers, their env vars, and how a custom OpenAI-compatible endpoint (self-hosted vLLM, a local proxy) plugs in.

The picker filters to providers you can actually use

/model shows only the providers you have a key configured for (see API keys), so abundance doesn't drown the list with providers you can't call anyway. Add a second provider's key and its models appear in the picker immediately, no restart required.

Each entry in the picker surfaces:

• Pricing — per-provider, per-model.
• Thinking depth, for models that support configurable reasoning effort.
• Per-conversation selection — switching models mid-conversation is a session-level choice, not a global one.

Refreshing the catalog

The cached models.dev fetch can be forced to refresh (dropping the on-disk cache first) rather than waiting for the next opportunistic update — useful right after a new model ships upstream and you don't want to wait.

Costs stay visible

Per-turn token counts and cost are shown in the footer continuously, and a context-usage gauge warns before you're at risk of hitting the window limit — both independent of which model or provider you're currently on.