caniollama

Does this local Ollama model actually support tool calling? Find out in seconds, not by reading a six-month-old blog post.

uvx caniollama check qwen2.5-coder:7b

or test everything you've pulled:

uvx caniollama check

What it does

Runs a small, fixed test suite against a locally running Ollama instance and classifies the result into one of three real answers:

1. Structured pass -- the model returns a proper tool_calls field.
2. JSON-in-content fallback -- the model knows tool calling exists but emits the intent as JSON-shaped text instead of the structured field.
3. Rejected by backend -- Ollama 400s the moment tools is included.

After a run, you'll be asked -- every time, explicitly -- whether to submit the result (metadata only: model name, quantization, Ollama version, pass/fail per case) to the public caniollama registry. See docs/privacy.md for exactly what is and isn't ever sent.

Install

No install needed for a one-off check:

uvx caniollama check <model>
# or
pipx run caniollama check <model>

For repeated use:

uv tool install caniollama

Development

uv sync
uv run pytest
uv run ruff check .
uv run mypy src

License

MIT