Baretools AI

Changelog

Zero runtime dependencies: pip install baretools-ai installs nothing but the package itself. Standard-library dataclasses work as tool parameter types out-of-the-box; Pydantic is opt-in via baretools-ai[pydantic].
Dataclass parameters: @dataclass fields are reflected into the JSON Schema and dict arguments from the LLM are coerced into the dataclass instance before invocation.
Pydantic model parameters: BaseModel subclasses are supported as parameter types. Schema comes from model_json_schema(); coercion uses model_validate().
Streaming: execute_stream() and execute_stream_async() yield ToolResult values as each call completes.
Multi-provider helpers: parse_tool_calls(response, provider=...) and format_tool_results(results, provider=...) for OpenAI, Anthropic, and Gemini.
Live provider examples: examples/openai_agent.py, examples/anthropic_agent.py, examples/gemini_agent.py — tested BMI agent loops for each provider.
Optional W&B Weave tracing: gated on WEAVE_PROJECT env var; zero overhead when unset.
Version from metadata: baretools.__version__ reads from importlib.metadata; pyproject.toml is the single source of truth.

Multi-provider schema support via ToolRegistry.get_schemas(provider=...).
Gemini output uses the functionDeclarations shape with OpenAPI subset compliance.
get_schemas() returns defensive copies.
strict=True for OpenAI structured-outputs mode.
Breaking: RegisteredTool.schema renamed to RegisteredTool.parameters.

This site is open source. Improve this page.