autoComplete

autoComplete is an open-source, local-first macOS inline autocomplete app. It is designed as a native menu-bar app that can either use an already-running local LLM server, such as Ollama or LM Studio, or a built-in MLX-based model path.

The first MVP focuses on:

• Cross-app text completions through macOS Accessibility APIs.
• Inline suggestions near the caret.
• Next-word acceptance with Tab and full acceptance with Option + Tab.
• OpenAI-compatible local servers, including Ollama and LM Studio.
• A preview built-in MLX provider for local model paths.
• Per-app policy defaults and a privacy-first settings model.

Typing history is not stored by default. Personalization storage is reserved for a future opt-in encrypted store.

Build

swift build
swift test

Run the development binary:

swift run autoComplete

Create a local .app bundle for manual testing:

./Scripts/package-app.sh
open dist/autoComplete.app

The app needs Accessibility permission to read focused text fields and insert accepted completions. Clipboard and screen context are off by default.

Local Server Providers

The default provider is Ollama:

• Base URL: http://localhost:11434/v1
• Model: llama3.2

LM Studio is also supported:

• Base URL: http://localhost:1234/v1

Any OpenAI-compatible local endpoint can be configured in Settings.

Built-in MLX Provider

The MLXProvider is currently a lightweight process-backed preview that calls python3 -m mlx_lm.generate against a local model path. It establishes the app interface for built-in inference without committing model weights to the repo. The production path should replace this runner with in-process mlx-swift-lm once the model catalog and license gating are finalized.

Backlog

• Opt-in encrypted writing history and personalization.
• Emoji suggestions.
• Typo indicator and full autocorrect.
• Screen-aware ranking and clipboard-aware prompts.
• Per-app instructions.
• Statistics.
• Full curated model catalog.
• Mid-line completion polish.