A complete AI coding IDE that runs entirely on your own machine. The model runs in-process; your code never leaves the device. No API key, no account, and no network dependency once the model is downloaded. Open a folder, load a model, and start pairing.
First run
- Install the app. On macOS drag it to Applications. Windows shows a SmartScreen
notice — More info → Run anyway. On Linux,
chmod +xthe AppImage. - macOS only: clear the quarantine flag. The build is ad‑hoc signed and not notarized —
notarization needs a paid Apple Developer account — so macOS says it “cannot verify this app is
free of malware” and refuses to open it. Press Done on that dialog (not “Move to Trash”),
then run this once:
Or: double-click the app, dismiss the warning, then open System Settings → Privacy & Security, scroll to Security, and press Open Anyway. Right-click → Open no longer works — Apple removed that bypass in macOS 15.
xattr -dr com.apple.quarantine "/Applications/Local Language Machine.app" - Open a folder. This becomes the workspace the assistant can read and edit.
- Load a model from the Assistant panel. The app ships without one — the default
Qwen2.5-Coder 7B is a 4.7 GB one-time download (1.0 GB and 0.1 GB options
are also available). It caches to
~/.local-language-machine, and nothing touches the network after that.
What you get
🔒 Private by designInference runs in-process. Files and prompts stay on disk, and the network switch in the status bar is off by default.
⚡ Fast turnsLive KV cache — each turn only evaluates new tokens.
🧩 Real agentReads, edits, greps, runs commands through a permissioned tool loop.
🎓 On-device tuningBuilt-in LoRA training and evaluation — MLX or Unsloth.
Or run from source
# requires Node ≥ 22 — GPU used automatically when present git clone https://github.com/siroxou/local-language-machine.git cd local-language-machine npm install npm run preview
Serves the same UI at localhost:7433. The inference engine is bundled — no
extra services.