ollmlx

Run local LLMs on Apple Silicon via mlx-lm. Menubar app + CLI with OpenAI-compatible and Ollama-compatible APIs on localhost:11434.

Requirements

• Apple Silicon Mac (M1 or later)
• macOS 14+
• Internet connection for first run (downloads uv, mlx-lm, and models from Hugging Face)

Installation

Homebrew (recommended)

brew tap darrylmorley/ollmlx
brew install --cask darrylmorley/ollmlx/ollmlx

Manual

1. Download the latest ollmlx-x.x.x.dmg from the releases page
2. Drag ollmlx.app to /Applications
3. Launch from Applications or Spotlight

First run

On first launch, ollmlx will:

1. Bootstrap Python environment — installs uv and creates a venv at ~/.ollmlx/venv with mlx-lm and huggingface-hub. A progress sheet shows the status. If something goes wrong, the error output is displayed with a Retry button. If a venv already exists at ~/.ollmlx/venv, this step is skipped automatically.

2. Offer to install the CLI — creates a symlink at /usr/local/bin/ollmlx so you can use ollmlx from any terminal. Requires your admin password. You can skip this and install later from Settings.

Usage

Menubar app

The daemon starts automatically when the app launches — no manual setup needed. Click the brain icon in your menu bar to:

• See the current server status (green = running, grey = stopped, red = error)
• Select and switch models from your cached models
• Start / stop the server
• Pull new models from Hugging Face
• Copy localhost:11434 to your clipboard
• Open server logs
• Access Settings
• Check for updates

CLI

The daemon is started automatically by the menubar app. If you installed via Homebrew, the ollmlx CLI is available immediately in any terminal. If you want to run CLI-only without the app, start the daemon manually first:

ollmlx serve

Then use any command from a separate terminal:

Command	Description
ollmlx serve	Start the daemon (control API on :11435, public API on :11434)
ollmlx run <model> [prompt]	Start a model and send a prompt, or enter interactive REPL
ollmlx pull <model>	Download a model from Hugging Face
ollmlx list	List locally cached models
ollmlx show <model>	Show model details (size, quantisation, context length)
ollmlx rm <model>	Remove a model from the local cache
ollmlx stop	Stop the currently running model
ollmlx ps	Show the running model (name, size, port)
ollmlx status	Print raw daemon status as JSON

Examples:

# Pull and run a model (start the app first, or run `ollmlx serve` in a terminal)
ollmlx pull mlx-community/Llama-3.2-3B-Instruct-4bit
ollmlx run mlx-community/Llama-3.2-3B-Instruct-4bit "What is the capital of France?"

# Interactive chat
ollmlx run mlx-community/Llama-3.2-3B-Instruct-4bit
>>> What is quantum computing?
>>> /bye

# List cached models
ollmlx list

# Check what is running
ollmlx ps

Ollama migration

ollmlx exposes both OpenAI-compatible (/v1/chat/completions, /v1/models) and Ollama-compatible (/api/tags, /api/chat, /api/generate) endpoints on the same default port as Ollama (localhost:11434). Any tool or client that works with Ollama should work with ollmlx without reconfiguration.

If an Ollama client requests a different model than the one currently running, ollmlx will automatically stop the current model and start the requested one.

Important: Ollama and ollmlx both use port 11434. Quit Ollama before launching ollmlx, or change one of the ports in Settings.

• Open WebUI — set the Ollama URL to http://localhost:11434. Model list and chat work natively via the Ollama API
• Continue, Cursor, or other IDE plugins — point them at http://localhost:11434 (usually the default)
• Python/JS clients — any OpenAI-compatible client library works, just set the base URL to http://localhost:11434/v1

Optional Ollama CLI shim

If you have scripts or tools that call the ollama binary directly, you can create a compatibility symlink from Settings → Compatibility → "Install Ollama shim". This creates /usr/local/bin/ollama pointing to /usr/local/bin/ollmlx.

If a real Ollama installation is detected at that path, the toggle will warn you and refuse to overwrite it.

Building from source

# Clone the repository
git clone https://github.com/darrylmorley/ollmlx.git
cd ollmlx

# Build the CLI only
swift build --target ollmlx

# Build all targets (including the app framework)
swift build

# Run tests
swift test

# Build the app via Xcode — open the project,
# select the OllmlxApp scheme, and hit Cmd+R
open ollmlx.xcodeproj

# Create a signed DMG (requires Developer ID)
bash Scripts/build_dmg.sh

Architecture

Sources/
  OllmlxCore/    Shared framework (no AppKit/SwiftUI) — server management,
                 proxy, control API, model store, config, logging
  OllmlxApp/     Menubar app (AppKit + SwiftUI) — status item, popover,
                 settings, bootstrap, Sparkle auto-update
  ollmlx/        CLI (ArgumentParser) — 9 commands against the control API

Port	Purpose
11434	Public OpenAI-compatible + Ollama-compatible API (clients connect here)
11435	Internal control API (CLI and app only)
Ephemeral	mlx_lm.server internal port (allocated dynamically)

Known limitations

• One model at a time — v1 supports running a single model. Switching models stops the current one first. External clients (e.g. Open WebUI) can trigger model switches automatically via /api/chat or /api/generate.
• Mac App Store not supported — ollmlx requires com.apple.security.cs.allow-unsigned-executable-memory to spawn the Python/MLX process and com.apple.security.temporary-exception.sbpl for process execution outside the sandbox. These entitlements are incompatible with the Mac App Store. Distribution is via signed and notarised DMG only.
• Apple Silicon only — mlx-lm requires Apple Silicon (M1+). Intel Macs are not supported.
• Ollama conflict — ollmlx and Ollama cannot run simultaneously as both default to port 11434. Either quit Ollama first or change the port in Settings.

License

MIT