System-wide inline autocomplete for macOS, powered by a language model that runs
entirely on your Mac. No cloud, no account, no telemetry.
Website ·
Download ·
Changelog
AutoSuggest watches the text field you're typing in, asks a local model what comes next, and shows the suggestion inline — press Tab to accept, Esc to dismiss. It works in any app, and your keystrokes never leave the machine.
- Private — inference, context, and personalization stay on-device. No accounts, no analytics.
- Fast — 100–300 ms suggestions via Ollama, llama.cpp, or CoreML; tuned for Apple Silicon.
- Yours to tune — per-app exclusion rules, PII filtering, battery-aware pause, and opt-in personalization that learns your style.
- Open — GPL v3, written in Swift 6.2 with strict concurrency.
Download (recommended). Grab the latest signed & notarized build from the
releases page, drag
AutoSuggest.app to /Applications, and open it. Or one line:
curl -fsSL https://raw.githubusercontent.com/2002Bishwajeet/autosuggest/main/scripts/install.sh | bashYou also need a local model runtime — the quickest is Ollama:
brew install ollama && ollama serve
ollama pull qwen2.5:1.5b- Grant Accessibility and Input Monitoring when prompted (System Settings → Privacy & Security). Both are required for a system-wide autocomplete; nothing works without them.
- Start typing in any text field. Suggestions appear inline — Tab/Enter to accept, Esc to dismiss.
AutoSuggest lives in the menu bar; click the ghost glyph to pause, switch models, exclude an app, or open settings.
Requirements: macOS 13 (Ventura)+, Apple Silicon recommended (Intel works, slower).
Pick any; the engine tries them in order and falls through automatically.
| Runtime | Setup |
|---|---|
| Ollama (recommended) | brew install ollama → ollama pull qwen2.5:1.5b |
| llama.cpp | llama-server -m model.gguf --port 8080 |
| CoreML | On-device via the Apple Neural Engine — point Settings → Model Source at a CoreML manifest |
Everything runs locally. Accepted suggestions are never logged; optional telemetry is off by default and content-free. Personalization is opt-in, PII-filtered, encrypted at rest, and never transmitted. AutoSuggest stays silent in password fields and macOS secure input. Read the privacy source — it's all auditable.
# Library + menu-bar runner (fast iteration)
swift build
swift run AutoSuggestRunner
swift test
# The real app target (correct for permission testing & distribution)
cd macos && xcodegen generate
open AutoSuggestDesktop.xcodeproj # scheme: AutoSuggestDesktop, Cmd+RUse the Xcode app target for anything permission-sensitive — it builds a real
bundled AutoSuggest.app with a stable bundle ID. See CLAUDE.md
for the architecture map and conventions.
Project layout
| Path | What |
|---|---|
Sources/AutoSuggestApp/ |
The library: input → context → policy → inference → overlay → insertion pipeline |
macos/ |
xcodegen spec + the distributable Xcode app shell |
website/ |
Marketing site (static, deployed to Cloudflare Pages) |
training/ |
Fine-tuning scripts (MLX / Colab) — see docs/FINE_TUNING.md |
docs/ |
Architecture, local setup, fine-tuning |
Config lives at ~/Library/Application Support/AutoSuggestApp/config.json
(runtime order, model source, exclusion rules); it's created on first run and
migrated forward across versions.
Train a small model on your own writing — on-device with MLX, or free on a Colab GPU — then import the GGUF with one command. See docs/FINE_TUNING.md.
Issues and PRs welcome. Run swift test and swiftformat Sources Tests --lint
before pushing; CI runs both. Touching the insertion, policy, or privacy paths?
Read the "Critical paths" section in CLAUDE.md first.