Agent Toolbelt Framework

A minimal Python framework for registering, managing, and invoking tool functions for LLM (Large Language Model) interactions and agentic AI.

Features

• Singleton tool registry for dynamic function registration and lookup
• Decorator for easy tool registration
• Automatic metadata generation for tool functions (for LLM tool use)
• Unified tool invocation interface with argument parsing
• Custom exception for missing tools
• Type hints throughout for clarity and static analysis
• Unit tests and linting included

Installation

1. Clone the repository:

   git clone https://github.com/simonharris/agent_toolbelt.git
   cd agent_toolbelt

2. Create a virtual environment and install dependencies:

   make install

Usage

Registering a Tool

from framework import register_tool

@register_tool('add')
def add(a: str, b: str) -> str:
    """Add two numbers as strings and return the sum as a string."""
    return str(int(a) + int(b))

Running a Tool

from framework import run_tool
import json

result = run_tool({
    "name": "add",
    "arguments": json.dumps({"a": "2", "b": "3"})
})
print(result)  # Output: '5'

Generating Tool Metadata

from framework import generate_tool_metadata, TOOL_REGISTRY
meta = generate_tool_metadata(TOOL_REGISTRY['add'])
print(meta)

Integrating with an LLM Call

You can use the generated tool metadata as part of your LLM API call (e.g., OpenAI, Anthropic, etc.).

# Example: Using with OpenAI's chat API
from framework import get_tools

tools = get_tools()  # List of tool metadata dicts

response = client.chat.completions.create(
  model=MODEL,
  messages=messages,
  tools=tools,  # Pass the tool metadata here
  tool_choice="auto"
)

This allows the LLM to see and use your registered tools automatically during the conversation.

Testing

Run all tests with:

make test

Linting and Code Quality

• Check PEP8/style issues:

  make lint

• See a code quality score:

  make pylint

License

MIT