macOS-wide autocomplete that runs a local LLM on every keystroke

What it does KeyType is a menu-bar utility that watches the focused text field in any macOS app, predicts a short continuation using a local LLM, and renders it as inline ghost text. Hit Tab to accept the suggestion; keep typing to ignore it.

The interesting bit Beneath the surface, the repo is sliced into nine local Swift packages that solve the unglamorous problems of system-wide autocomplete: scraping the caret position via accessibility APIs, budgeted prompt assembly, constrained generation with logit masking, and per-app compatibility overrides. It is engineered like a research artifact, not a thin wrapper around llama.cpp.

Key highlights

• Runs fully offline; no keystrokes leave the machine
• Uses macOS accessibility APIs to snapshot focused text fields across apps
• Renders ghost-text overlays and inserts completions via pasteboard or synthetic keystrokes
• Includes constrained generation logic (trie admissibility, logit masking) to keep predictions in bounds
• MIT-licensed alternative to the closed-source app Cotypist

Verdict macOS developers curious about on-device LLM inference and accessibility-driven UI automation should dig into the package layout. If you just want polished autocomplete today, the README frames this as an alternative to the closed-source original.