llm-models

# LLM Models Query LLM model information from [models.dev](https://models.dev). The cached data includes model IDs, display names, pricing (per 1M tokens), context/output limits, modalities, and capability flags (reasoning, tool_call, etc.) across 90+ providers. ## Script `scripts/llm-models.py` — requires `uv` to run. All commands below assume the skill directory as working directory. ## Commands ```bash # Refresh cache (auto-runs on first use, re-run to get latest data) uv run scripts/llm-models.py update # List all providers uv run scripts/llm-models.py providers # List models — query format is provider:pattern (pattern is optional regex, case-insensitive) uv run scripts/llm-models.py list anthropic # all Anthropic models uv run scripts/llm-models.py list anthropic:opus openai:gpt-4 # multiple queries at once uv run scripts/llm-models.py list -p anthropic:4-6 # with pricing (USD per 1M tokens) ``` ## Detailed data in cache The script outputs model IDs and optionally pricing, but the cache file (`~/.local/share/llm-models/models.json`) contains richer per-model data. When the user needs details beyond IDs and pricing — such as context window size, output token limit, supported modalities, or whether a model supports reasoning — read the cache file directly and extract the relevant fields. Each model entry has this structure: - `id` — exact model identifier to use in API calls - `name` — human-friendly display name - `family` — model family (e.g. "claude", "gpt", "gemini") - `cost.input` / `cost.output` — USD per 1M tokens - `limit.context` / `limit.output` — token limits - `modalities.input` / `modalities.output` — e.g. ["text", "image"] - `reasoning`, `tool_call`, `attachment` — boolean capability flags - `open_weights` — whether the model weights are open - `release_date`, `last_updated` ## Workflow 1. **Run the script** from the skill directory using `uv run scripts/llm-models.py`. Always include `-p` when pricing is relevant to the question. 2. **For richer queries** (context windows, capabilities, comparisons across multiple fields), read the cache JSON directly and filter/format the results for the user. 3. **Present results concisely** — the user usually wants a quick answer, not a wall of data. If they ask "what's the model ID for Claude Opus?", give them the ID. If they ask to compare pricing, show a short table. 4. **If data seems stale or a model is missing**, run `update` first, then retry.