crowact 0.1.0

A lightweight ReAct agent framework with tool calling, streaming chunks, and pluggable providers.

CrowAct

crowact is a lightweight ReAct agent framework with tool calling, streaming chunks, and pluggable providers.

Distribution name: crowact
Python import name: CrowAct

Requirements

• Python 3.10+
• requests
• python-dotenv

Install

pip install crowact

For local development in this repository:

pip install -e .

Quick Start

from CrowAct import Agent, LLMProvider
from CrowAct.agent.tools import get_tools

TOOLS = get_tools("test_folder")


def main() -> None:
    provider = LLMProvider.from_anthropic_env("byte2.env")
    agent = Agent(
        provider=provider,
        model="deepseek-v3.2",
        system_prompt=(
            "You are a concise ReAct-style assistant. "
            "Use tools first when needed. "
            "You may call tools in multiple steps. "
            "After receiving tool results, provide the final answer."
        ),
        tools=TOOLS,
    )

    question = (
        "Please use tools to calculate this step by step: "
        "first compute 1+19, then take the square root of the result, "
        "and finally add 100."
    )

    for chunk in agent.run(question, stream=True):
        print(chunk)

    print(agent.last_answer)


if __name__ == "__main__":
    main()

Tool Files

get_tools(folder) loads every *.py file in the target folder and registers functions marked with the @tool(...) decorator.

Example folder:

test_folder/
  tool1.py
  tool2.py

Example tool:

from CrowAct.agent.tools import tool


@tool(
    description="Compute the square root of a number.",
    param_descriptions={"number": "The number to take the square root of"},
)
def sqrt_tool(number: float) -> float:
    return number ** 0.5

Agent Output

Agent.run(...) always returns an iterator of chunks.

Chunk types:

• {"type": "text", "text": "..."}
• {"type": "tool_use", "id": "...", "name": "...", "input": {...}}
• {"type": "tool_result", "tool_use_id": "...", "content": "..."}

Behavior:

• stream=True: text is yielded incrementally as streaming chunks.
• stream=False: text is yielded as complete blocks, but the API remains iterator-based.

The final plain-text answer is stored in agent.last_answer.

History

You can pass prior conversation context through history.

agent = Agent(
    provider=provider,
    model="deepseek-v3.2",
    system_prompt="You are helpful.",
    history=[
        {"role": "user", "content": [{"type": "text", "text": "Previous question"}]},
        {"role": "assistant", "content": [{"type": "text", "text": "Previous answer"}]},
    ],
)

history is prepended to the current run, but it is not mutated after execution.

Provider Setup

Anthropic-style

provider = LLMProvider.from_anthropic_env(".env")

Expected environment variables:

ANTHROPIC_BASE_URL=https://your-endpoint
ANTHROPIC_API_KEY=your-api-key

Endpoint pattern:

• .../v1/messages

OpenAI-style

provider = LLMProvider.from_openai_env(".env")

Expected environment variables:

OPENAI_BASE_URL=https://your-endpoint
OPENAI_API_KEY=your-api-key

Endpoint pattern:

• .../chat/completions

Current Limitation

LLMProvider supports both anthropic and openai endpoint styles, but the request body and tool-calling flow are currently implemented around the Anthropic-style message format. If you use an OpenAI-compatible endpoint, confirm that it accepts the same request shape before relying on it.

Local Example

This repository includes:

• main.py
• test_folder/tool1.py
• test_folder/tool2.py

License

Add a license before publishing to PyPI.