MLX (Apple Silicon)

MLX is Apple's machine learning framework optimized for Apple Silicon. Combined with mlx-lm, it can serve local models with an OpenAI-compatible API that AgentXchain connects to via api_proxy.

Which adapter?

api_proxy with provider: "ollama" (OpenAI-compatible) — MLX's server exposes the same API format.

Prerequisites

• Apple Silicon Mac (M1/M2/M3/M4)
• Python 3.10+ with mlx-lm installed (pip install mlx-lm)
• A model downloaded (e.g., mlx-community/Qwen3-Coder-30B-A3B-4bit)
• agentxchain CLI installed

Start the MLX server

mlx_lm.server --model mlx-community/Qwen3-Coder-30B-A3B-4bit --port 8080

This serves an OpenAI-compatible API at http://localhost:8080/v1/chat/completions.

Configuration

{
  "runtimes": {
    "mlx-dev": {
      "type": "api_proxy",
      "provider": "ollama",
      "model": "mlx-community/Qwen3-Coder-30B-A3B-4bit",
      "auth_env": "MLX_API_KEY",
      "base_url": "http://localhost:8080/v1/chat/completions"
    }
  },
  "roles": {
    "dev": {
      "runtime": "mlx-dev",
      "mandate": "Implement features and fix bugs",
      "authority": "proposed"
    }
  }
}

Set a dummy API key (MLX server doesn't require auth):

export MLX_API_KEY="mlx"

Verify the connection

mlx_lm.server --model mlx-community/Qwen3-Coder-30B-A3B-4bit --port 8080 &
export MLX_API_KEY="mlx"
agentxchain connector check

Recommended models

Model	Params	Unified Memory
mlx-community/Qwen3-Coder-30B-A3B-4bit	30B (3B active)	~8GB
mlx-community/deepseek-coder-v3-16b-4bit	16B	~10GB
mlx-community/codestral-22B-v0.1-4bit	22B	~14GB

Gotchas

• Apple Silicon only: MLX does not run on Intel Macs or Linux.
• Unified memory: Models share memory with the rest of the system. Leave headroom for your IDE and other tools.
• Quantization: 4-bit quantized models are the practical choice for most Apple Silicon Macs. Full-precision models require significantly more memory.
• Provider field: Use "ollama" as the provider since MLX's server implements the OpenAI-compatible API format. The base_url override points to the actual MLX endpoint.