Models & providers

Lime ships four provider backends. Each maps onto its own Cargo crate and its own wire protocol; the runtime selects the right backend for the active model.

OpenAI Responses API, WebSocket streaming, ZDR-encrypted reasoning items.

Anthropic Messages API, adaptive thinking, tiered prompt caching.

Google Gemini Generative Language API with streaming, thinking on Gemini 2.5+.

OpenAI-compatible Groq, OpenRouter, Together, Mistral, Ollama, vLLM, and friends.

Custom models Register your own model in ~/.lime/models.json.

How model selection works

The active model is resolved in this order:

1. --model SLUG command-line flag.
2. LIME_MODEL environment variable.
3. models.default.slug in merged config.
4. The hard-coded default gpt-5.4.

The slug is looked up in the merged catalog — built-in entries plus any user entries in ~/.lime/models.json. User entries override built-ins on slug collision; new slugs are appended.

lime --model claude-opus-4-7
lime --model o4-mini --reasoning-effort high
lime --model pro                    # alias for gemini-2.5-pro
lime --model llama-3.3-70b          # a custom slug from ~/.lime/models.json

Inside the REPL, /model shows the merged catalog and switches mid-session:

/model claude-sonnet-4-6
/reasoning xhigh           # only on models that support it
/verbosity low

Per-model capabilities

The catalog records what each model supports. Capabilities flagged on a model entry:

Capability	What it gates
supports_vision	The model accepts images.
supports_parallel_tool_calls	Multiple tool calls per turn.
supports_reasoning_summaries	OpenAI-style reasoning summaries (encrypted_content).
supports_extended_thinking	Anthropic-style extended thinking.
param_preset	Which --reasoning-effort / --verbosity values the provider will accept.

param_preset values:

• none — neither reasoning effort nor verbosity is applicable.
• verbosity_only — verbosity is forwarded; reasoning effort is ignored.
• reasoning_only — reasoning effort is forwarded; verbosity is ignored.
• reasoning_and_verbosity — both are forwarded.

Bring your own catalog

~/.lime/models.json is merged with the built-in catalog. You can override a built-in (e.g. raise an auto-compact threshold) or register a brand-new slug.

lime model add \
  --slug llama-3.3-70b \
  --provider openai-compat --service groq \
  --api-id llama-3.3-70b-versatile \
  --context-window 128000

lime model list           # show the merged catalog with built-in/user tags
lime model rm llama-3.3-70b

Set LIME_MODELS_FILE to override the path (primarily for testing). If the file is absent or unparseable, Lime silently falls back to the bundled catalog. See Custom models for the full schema.