agent-framework-mlx 0.6.0

MLX integration for the Agent Framework

Agent Framework MLX

This library provides an Apple MLX LM backend for the Agent Framework. It allows you to run Language Models locally on macOS with Apple Silicon using the mlx-lm library.

Features

• Local Inference: Run models directly on your Mac using Apple Silicon
• Tool Calling: Native support for agentic tool invocation.
• Observability: Built-in OpenTelemetry tracing, metrics, and token usage tracking.
• Streaming Support: Full support for streaming responses.
• Configurable Generation: Fine-tune generation parameters like temperature, top-p, repetition penalty, and more.
• Message Preprocessing: Hook into the pipeline to modify messages before they are converted to prompts.
• Agent Framework Integration: Seamlessly plugs into the Agent Framework's BaseChatClient interface.

Installation

Ensure you have Python 3.9+ and are running on macOS with Apple Silicon.

pip install agent-framework-mlx

Or install from source:

git clone https://github.com/filipw/agent-framework-mlx.git
cd agent-framework-mlx
pip install -e .

Usage

Basic Example

import asyncio
from agent_framework import ChatMessage, Role, ChatOptions
from agent_framework_mlx import MLXChatClient, MLXGenerationConfig

# Initialize the client
client = MLXChatClient(
    model_path="mlx-community/Phi-4-mini-instruct-4bit",
    generation_config=MLXGenerationConfig(
        temp=0.7,
        max_tokens=500
    )
)

# Create messages
messages = [
    ChatMessage(role=Role.SYSTEM, text="You are a helpful assistant."),
    ChatMessage(role=Role.USER, text="Why is the sky blue?")
]

# Get response
response = await client.get_response(messages=messages)
print(response.text)

Agent with Tools

You can easily create agents capable of calling tools (Python functions) using the client:

from typing import Annotated

def calculate_bmi(weight: float, height: float) -> str:
    """Calculates BMI."""
    return f"{weight / (height ** 2):.2f}"

# Create an agent with the tool
agent = client.as_agent(
    name="HealthAssistant",
    instructions="You are a helpful assistant.",
    tools=[calculate_bmi]
)

response = await agent.run("Calculate BMI for 70kg and 1.75m")
print(response)

Workflow Integration

You can use the client as backbone for Agent Framework agents when building agentic workflows:

from agent_framework import ChatAgent

# notice the client constructed in the previous example now backs the local agent
local_agent = client.as_agent(
    name="Local_MLX",
    instructions="You are a helpful assistant."
)

remote_agent = ChatAgent(
    name="Cloud_LLM",
    instructions="You are a fallback expert. The previous assistant was unsure. Provide a complete answer.",
    chat_client=azure_client
)

builder = WorkflowBuilder()
builder.set_start_executor(local_agent)

builder.add_edge(
    source=local_agent,
    target=remote_agent,
    condition=should_fallback_to_cloud
)

workflow = builder.build()

Streaming

async for update in client.get_streaming_response(messages=messages):
    print(update.text, end="", flush=True)

Configuration

You can configure the client using environment variables or a .env file. Using environment variables like MLX_MODEL_PATH allows you to omit arguments in code.

export MLX_MODEL_PATH="mlx-community/Phi-4-mini-instruct-4bit"

# No arguments needed if env vars are set
client = MLXChatClient()

Advanced Configuration

You can configure generation parameters globally via MLXGenerationConfig or per-request via MLXChatOptions.

config = MLXGenerationConfig(
    temp=0.7,
    top_p=0.9,
    repetition_penalty=1.1,
    seed=42
)

Message Preprocessing

You can intercept and modify messages before they are sent to the model. This is useful for injecting instructions or formatting content.

def inject_instruction(messages):
    if messages:
        messages[-1]["content"] += "\nIMPORTANT: Be concise."
    return messages

client = MLXChatClient(
    model_path="...",
    message_preprocessor=inject_instruction
)

Requirements

• macOS
• Apple Silicon
• Python 3.9+

License

This project is licensed under the MIT License. See the LICENSE file for details.