MLX Control

English | 한국어

A native macOS menu bar app to manage your local mlx-lm inference server — start, stop, monitor resources, and download models, all without touching the terminal.

Table of Contents

• Features
• Requirements
• Installation
• Usage
• Logs
• Notarization
• Security
• License

Features

• Server control — Start / Stop / Restart mlx_lm.server from the menu bar
• Real-time monitoring — MLX process RAM & CPU (accurate) + system GPU % / GPU mem (via ioreg, no sudo)
• Mini sparkline graph — GPU utilization history
• Warm-up — Pre-load the model into memory + measure tokens/sec
• Model management — Auto-detect installed MLX models from HuggingFace cache
• Model search — Search mlx-community on HuggingFace with size, description, and README
• Model download — Download any mlx-community model directly from the app
• Model delete — Remove a cached model with disk space preview
• Resource alerts — macOS notifications when RAM or free memory crosses a threshold
• Copy endpoint — One-click copy http://127.0.0.1:8080/v1 for use in other tools
• Login item toggle — Auto-launch at login
• Move to /Applications — Prompted on first launch if running from elsewhere

Requirements

Requirement	Detail
macOS	14.0 (Sonoma) or later
Hardware	Apple Silicon (M-series) — GPU monitoring via Metal/ioreg
mlx-lm	pip install mlx-lm or uv tool install mlx-lm
Models	At least one MLX model in ~/.cache/huggingface/hub/

Installation

Option A — Build from source (recommended, no Gatekeeper warning)

git clone https://github.com/wonsss/MLXControl.git
cd MLXControl
./build.sh          # builds + installs to /Applications
open /Applications/MLXControl.app

Requires Xcode Command Line Tools: xcode-select --install

Option B — Download pre-built (GitHub Releases)

1. Download MLXControl-x.x.x.dmg from Releases
2. Open the DMG → drag MLXControl.app to /Applications
3. Notarized — opens without Gatekeeper warnings.

Option C — Homebrew

brew tap wonsss/tools
brew install --cask mlxcontrol

Usage

1. Install mlx-lm and download at least one model:

   pip install mlx-lm
   python -m mlx_lm.manage --scan   # see what's cached

2. Click ⚡ in the menu bar
3. Select a model → Start
4. Once status shows Up, click 🔥 Warm to pre-load the model
5. Point your AI agent (Hermes, Claude Code, Cursor, etc.) at http://127.0.0.1:8080/v1

Logs

Server logs are written to ~/Library/Logs/MLXControl/mlx_server.log. Click 📄 in the app to open them in Console.app.

Notarization

For distributors with an Apple Developer Program membership and a "Developer ID Application" certificate:

# one-time: store Apple ID + app-specific password in the local Keychain
./notarize.sh --store-credentials

# build, sign, submit to Apple, and staple
./notarize.sh

This produces MLXControl-notarized.zip ready for GitHub Releases. Credentials are stored in the local Keychain (never on the command line or in git). Generate an app-specific password at appleid.apple.com → Sign-In & Security → App-Specific Passwords. The Team ID is auto-detected from your signing certificate.

Security

• No shell string interpolation — all subprocess calls use Process with argument arrays
• Path validation — model delete validates the repo ID format and confines removal to the HuggingFace cache directory
• Tool auto-detection — mlx_lm.server and hf are located via PATH search, not hardcoded paths
• No network access except the HuggingFace API (search/metadata) and the local inference server

License

MIT — see LICENSE