Openclaw Mlx Setup

OpenClaw + MLX Local Model Setup

One-click setup for running local LLMs on Apple Silicon Mac with OpenClaw

[](https://support.apple.com/en-us/116943) [](LICENSE)

What is this?

This skill automatically sets up a local LLM on your Apple Silicon Mac and connects it to [OpenClaw](https://github.com/nichochar/openclaw). It handles everything:

1. Detects your Mac's RAM and selects the optimal model
2. Installs dependencies (mlx-lm)
3. Downloads the model from HuggingFace
4. Starts mlx_lm.server (OpenAI-compatible API)
5. Configures openclaw.json automatically
6. Tests the connection

Quick Start

# Clone and run
git clone https://github.com/jkf87/openclaw-mlx-setup.git
cd openclaw-mlx-setup
bash scripts/auto_install.sh

That's it. Your OpenClaw agents (Telegram bots, etc.) can now use local models.

Requirements

Item	Requirement
Mac	Apple Silicon (M1/M2/M3/M4)
RAM	8GB minimum
Python	3.10+
OpenClaw	Installed and configured

Auto-Selected Models by RAM

RAM	Model	Size	Use Case
8GB	Qwen3-4B-4bit	~2.5GB	Light chat
16GB	Qwen3-8B-4bit	~4.9GB	Best value
32GB	Qwen3-14B-4bit	~8.5GB	Balanced
64GB+	Qwen3-32B-4bit	~19GB	High performance

Architecture

[HuggingFace MLX Model] --> [mlx_lm.server :8080] --> [OpenClaw] --> [Telegram Bot / Agent]
                             OpenAI-compatible API     mlx provider

Options

# Auto-select model (default)
bash scripts/auto_install.sh

# Specify a model
bash scripts/auto_install.sh --model mlx-community/Qwen3-14B-4bit

# Change port
bash scripts/auto_install.sh --port 8081

# Restart without re-downloading
bash scripts/auto_install.sh --skip-download

Mac Chip Performance (8B Q4)

| Chip | Token/s (approx)