Providers and models¶

Register providers in init.lua so you can run plain smelt with no flags. For a first-time setup see Getting started.

smelt.provider.register("ollama", {
  type = "openai-compatible",
  api_base = "http://localhost:11434/v1",
  models = { "qwen3.6:27b" },
})

smelt.provider.register("openai", {
  type = "openai",
  api_base = "https://api.openai.com/v1",
  api_key_env = "OPENAI_API_KEY",
  models = { "gpt-5.5" },
})

Registering the same name again replaces the earlier entry.

Provider types¶

Choosing a model¶

On launch, smelt picks the model in this order:

1. --model
2. The model you last chose with /model
3. smelt.defaults.set({ model = "..." })
4. The first available model from your providers

Switch at any time with /model. When two providers expose the same model name, use provider/model.

Mode and reasoning¶

Set the starting mode and reasoning effort with the --mode, --reasoning-effort and related flags, or in init.lua. At runtime, Shift+Tab cycles modes and Ctrl+T cycles reasoning.

Reasoning effort sets how long the model thinks before answering. /reasoning opens a picker of the levels the active model supports. If you switch to a model that does not support the current level, smelt moves to that model's default and tells you.

Defaults and last-used picks¶

Model, mode and reasoning effort each have two layers: a default you set in init.lua, and the last value you picked in the TUI, which smelt restores on the next launch. The order is CLI flag, then last-used, then default. Resuming a session (--resume) uses that session's own values instead.

Pin defaults, and opt a value out of last-used recall:

smelt.defaults.set({
  model = "openai/gpt-5.5",
  mode = "plan",
  reasoning_effort = "high",
})

smelt.remember.set({
  mode = false,             -- always start in the default mode
  reasoning_effort = false, -- always start at the default effort
})

Helper models¶

Background features can use a cheaper model than your main session: title, compact, predict, btw and web_fetch. When unset they use the main model.

smelt.model.preferred("title", "openai/gpt-5-mini")
smelt.model.preferred("compact", "anthropic/claude-haiku-4-5")
smelt.model.preferred("predict", "openai/gpt-5-mini")

The model must belong to a registered provider and is resolved like /model. Plugins can define their own names.

Per-model options¶

A model can be a plain string or a table with overrides:

smelt.provider.register("ollama", {
  type = "openai-compatible",
  api_base = "http://localhost:11434/v1",
  models = {
    { name = "qwen3.6:27b", temperature = 0.8, top_p = 0.95, top_k = 40 },
    { name = "custom-model", input_cost = 2.0, output_cost = 8.0 },
  },
})

Custom reasoning models¶

For an endpoint that does not advertise reasoning support, declare it yourself, using the levels your server actually accepts:

smelt.provider.register("box", {
  type = "openai-compatible",
  api_base = "http://localhost:8080/v1",
  models = {
    {
      name = "Qwen3.8-27B",
      supports_reasoning = true,
      supported_reasoning_efforts = { "off", "low", "medium", "xhigh" },
      default_reasoning_effort = "xhigh",
    },
  },
})

With supports_reasoning = true, /reasoning off sends reasoning_effort = "none". /thinking off only hides thinking blocks; it does not turn reasoning off.

Pricing¶

Cost tracking is built in for popular models. Subscription providers (Codex, Copilot, Kimi Code) show zero cost. The session cost appears in the status bar and the running total in /stats. For other models, set the *_cost fields above; unknown models count as free.