mlx-community

Desk

Your Mac. Your models. Your files.

Desk is a native macOS assistant that runs MLX from Hugging Face on Apple Silicon — full conversations, optional live web research. Everything stays local; inference never needs an API key.

Apple Silicon mlx-vlm & mlx-lm Private by default Search · Thinking · Loop

Capabilities

Local MLX inference

Download Qwen 3.5, Qwen 3, Llama, Mistral, Gemma, and more from mlx-community into ~/.desk/models/. One-click download, switch, and delete to reclaim space. Runtime installs automatically on first launch.

Research-grade search

Enable Search per chat to query many engines, crawl pages into a local index, and answer with fresh evidence — news, docs, prices, weather — without vendor search APIs.

Filesystem tools

Create project folders, write READMEs, list directories, and read files under ~/Documents and the rest of your home folder. Protected paths stay off limits.

Agent modes

Thinking for harder reasoning, Loop for multi-step tasks with completeness checks, independent chats with memory — tuned per session, not one global setting.

Built for trust

Chats in SQLite on your Mac. Host binds to 127.0.0.1 only. Voice input in the composer. No cloud relay required to talk to your model.

Get Desk

This page is the project home — not the app. Build the Mac bundle from source:

source "$HOME/.cargo/env"
cd apps/desktop
npm install
npm run build:mac
open src-tauri/target/release/bundle/macos/Desk.app

Then Settings → Download model → New chat. On 8 GB RAM, start with Qwen 3.5 4B or 2B and Thinking off until you know your headroom.