baretools-ai 0.4.4

Minimal tooling primitives for LLM tool calling

Baretools AI

The un-framework for AI Engineers — Build AI Agents, Not Framework Wrappers.

Status: Alpha. Documentation  ·  API Reference  ·  Why Baretools?  ·  Changelog

---

Why

Modern agent frameworks own your prompts, your orchestration, and your state. That trade-off is fine for demos and POCs, but in production it costs you the control needed to build your own Agent Harness. You also end up knowing the framework better than the underlying engineering — which is exactly backwards.

Baretools handles only the mechanical glue between your Python functions and the LLM:

• Function → provider tool schema (OpenAI, Anthropic, Gemini, generic JSON Schema)
• Parsing tool calls out of provider responses
• Validating and executing those calls (sync, async, parallel, streaming)
• Formatting results back into provider-shaped messages

Everything else — prompts, loops, retries, memory, guardrails — stays in your code.

Install

pip install baretools-ai

Pre-requisite:

• Python >=3.10x
• Optional Pydantic support: pip install "baretools-ai[pydantic]".

Quickstart

from baretools import tool, ToolRegistry, parse_tool_calls, format_tool_results
from openai import OpenAI

@tool
def get_weather(location: str) -> str:
    """Get current weather for a location."""
    return f"Sunny, 72°F in {location}"

tools = ToolRegistry()
tools.register(get_weather)

client = OpenAI()
messages = [{"role": "user", "content": "What's the weather in Paris?"}]

max_iterations = 5 # Can be very high in real world agents
iteration = 0
while iteration < max_iterations:
    response = client.chat.completions.create(
        model="gpt-4.1",
        messages=messages,
        tools=tools.get_schemas("openai"),
    )
    iteration += 1
    message = response.choices[0].message
    messages.append(message)

    tool_calls = parse_tool_calls(message, "openai")
    if not tool_calls:
        print("Final Response:", message.content)
        break

    results = tools.execute(tool_calls)
    messages.extend(format_tool_results(results, "openai"))

You write the loop. Baretools handles the schema, parsing, execution, and formatting on each side.

Runnable Examples

Working agents for each provider are in examples/:

OPENAI_API_KEY=...    uv run python examples/openai_agent.py
ANTHROPIC_API_KEY=... uv run python examples/anthropic_agent.py
GOOGLE_API_KEY=...    uv run python examples/gemini_agent.py

Features

• Zero runtime dependencies — stdlib only; no transitive supply chain to audit
• Multi-provider schemas — tools.get_schemas("openai" | "anthropic" | "gemini" | "json_schema")
• Sync, async, streaming — execute, execute_async, execute_stream, execute_stream_async
• Parallel tool execution — pass parallel=True with max_workers (sync) or max_concurrency (async) to fan out independent calls
• Tool call hooks — before_tool and after_tool callbacks let you plug in guardrails, redaction, auditing, or tracing around every call
• Retries with structured events — pass on_event=... to observe attempts/retries/failures
• Type-driven validation — dataclasses work out of the box; pydantic BaseModels supported when installed
• Provider-native parsing/formatting — parse_tool_calls() and format_tool_results() for all four providers

See the docs site for the full API reference, advanced patterns, and design notes.

Development

uv sync --group dev
uv run ruff check .
uv run pytest -q

Contributing

Baretools is designed to be strictly minimal. Preferably, we only add features that are universally needed across tool-calling applications and difficult for developers to implement themselves. Please open an issue to discuss your idea before proposing a new feature.

License

MIT — see LICENSE.