System-wide inline autocomplete, powered by a language model that runs entirely on your machine. Private, fast, open source.
Requires macOS 13+
· Apple Silicon recommended · GPL v3
Everything runs locally on your Mac. No data leaves your device. No accounts, no cloud, no telemetry.
Suggestions in 100–300ms via Ollama, llama.cpp, or CoreML. Optimized for Apple Silicon.
GPL v3 licensed. Read every line, fork it, extend it. Built in Swift 6.2 with strict concurrency.
Per-app exclusion rules, PII filtering, battery-aware pause, and personalization that learns your style.
Install Ollama with Homebrew and pull a small model. That's the whole setup — about a minute.
It lives in your menu bar. Grant Accessibility and Input Monitoring when prompted.
Suggestions appear inline in any text field. Press Tab to accept, Esc to dismiss.
Easiest setup. Manages models for you. Great performance on Apple Silicon.
brew install ollama
ollama pull qwen2.5:1.5b
Direct GGUF model loading. Maximum control over quantization and parameters.
brew install llama.cpp
llama-server -m model.gguf
Apple Neural Engine inference with a bundled model. No server to run.
// Settings > Model Source
// Point to a CoreML manifest
No cloud. No telemetry. No accounts.
Fine-tune directly on Apple Silicon using MLX. No GPU rental needed — your M-series chip handles it.
pip install -r requirements-mlx.txt
python finetune_mlx.py --data-dir ./data
Upload your data to Google Colab and train with Unsloth on a free T4 GPU. Ready-made notebook included.
Open finetune_colab.ipynb in Colab
Upload data → Train → Download GGUF
Export the fine-tuned model as GGUF with an auto-generated Ollama Modelfile. One command to import.
ollama create my-autosuggest -f Modelfile
Training data is opt-in, PII-filtered twice, stored locally, and never transmitted.
Settings > Privacy > Enable collection
No. Suggestions are generated by a model running on your Mac — your text, the surrounding context, and any personalization never leave the device. The only network activity is what you set up yourself: downloading a model, or talking to a local runtime on 127.0.0.1.
Accessibility lets AutoSuggest read the text field you're focused on and insert a suggestion when you accept it. Input Monitoring lets it notice typing and the Tab / Esc shortcuts. A system-wide autocomplete can't work without them.
Not unless you turn on personalization. When you do, entries are PII-filtered, encrypted at rest, and kept on your Mac — never uploaded.
No. AutoSuggest detects secure text fields and macOS secure input mode and stays silent there. Common password-manager apps are excluded by default.
macOS 13 (Ventura) or later. Apple Silicon is recommended for sub-second suggestions; Intel Macs work but are slower.
Quit AutoSuggest from the menu bar, drag it out of /Applications, and revoke its permissions under System Settings → Privacy & Security. No background remnants.
Free, open source, and runs entirely on your machine. No account needed. No internet required.
$ curl -fsSL .../install.sh | bash