minimal-harness 0.2.3.post1

An exploration of making an agent sdk as lean as possible while being effective.

minimal-harness

Documentation: /docs

A lightweight Python agent harness for building LLM-powered agents with tool-calling support.

Latest version: 0.2.3

What This Project Is For

Minimal-harness is a lean framework for building agents that can call tools. It provides:

• OpenAI-compatible API - Works with any OpenAI-compatible API provider
• Tool system - Create tools via decorators; includes built-in tools (bash, file ops)
• AsyncIterator events - Real-time async iteration for chunks, tool start/end, execution events
• Conversation memory - Tracks token usage across interactions
• ESC stop support - Gracefully stop LLM streaming and tool execution

Architecture

The framework uses an event-driven architecture with AsyncIterator-based event handling:

Agent (OpenAIAgent) → Internal Events → FrameworkClient → Client-Facing Events

Event flow:

async for event in agent.run(user_input=[{"type": "text", "text": "..."}]):
    if isinstance(event, LLMChunk):
        # handle chunk
    elif isinstance(event, ToolEnd):
        # handle tool result

How to Build an App

Project Structure

A typical app looks like this:

my-app/
├── cli.py          # Entry point
└── tools.py        # Your custom tools

1. Create Your Entry Point

import argparse
from minimal_harness import OpenAIAgent
from minimal_harness.client import FrameworkClient
from minimal_harness.client.events import (
    AgentStartEvent,
    AgentEndEvent,
    LLMChunkEvent,
    ToolStartEvent,
    ToolEndEvent,
)

def main():
    parser = argparse.ArgumentParser(description="My AI agent")
    parser.add_argument("--base-url", required=True)
    parser.add_argument("--api-key", required=True)
    parser.add_argument("--model", default="qwen3.5-27b")
    args = parser.parse_args()

    agent = OpenAIAgent(
        api_key=args.api_key,
        base_url=args.base_url,
        model=args.model,
    )
    client = FrameworkClient(agent)

    async def run():
        async for event in client.run(user_input=[{"type": "text", "text": "What files are in the current directory?"}]):
            if isinstance(event, AgentStartEvent):
                print(f"Agent starting...")
            elif isinstance(event, LLMChunkEvent):
                print(event.content, end="", flush=True)
            elif isinstance(event, ToolStartEvent):
                print(f"\n[Calling tool: {event.name}]")
            elif isinstance(event, ToolEndEvent):
                print(f"\n[Tool result: {event.result[:100]}...]")

    import asyncio
    asyncio.run(run())

if __name__ == "__main__":
    main()

2. Add Custom Tools

Use the @register_tool decorator to add your own tools:

from minimal_harness.tool.registration import register_tool

@register_tool(
    name="get_weather",
    description="Get weather for a location",
    parameters={
        "type": "object",
        "properties": {"location": {"type": "string"}},
        "required": ["location"],
    },
)
async def get_weather(location: str) -> str:
    return f"The weather in {location} is sunny."

The decorator auto-registers the tool. Just import it before running the agent.

3. Run

python cli.py --base-url https://api.openai.com/v1 --api-key sk-... --model gpt-4o

Or set environment variables:

export MH_BASE_URL=https://api.openai.com/v1
export MH_API_KEY=sk-...
export MH_MODEL=gpt-4o
python cli.py

Built-in Tools

Tool	Description
bash	Execute shell commands with timeout
read_file	Read file contents with line range
create_file	Create new files
patch_file	Patch files (append, prepend, etc.)
delete_file	Delete files

Event Types

Event	Description
AgentStartEvent	Agent execution started
AgentEndEvent	Agent execution completed
LLMStartEvent	LLM generation started
LLMChunkEvent	LLM output chunk received
LLMEndEvent	LLM generation completed
ExecutionStartEvent	Tool execution started
ExecutionEndEvent	Tool execution completed
ToolStartEvent	Tool call started
ToolProgressEvent	Tool intermediate progress
ToolEndEvent	Tool call completed with result

Environment Variables

Variable	Description
MH_BASE_URL	API base URL
MH_API_KEY	API key
MH_MODEL	Model name (default: qwen3.5-27b)

Stop Mechanism

Press ESC during execution to gracefully stop LLM streaming and tool execution.