A local LLM that ghosts your cursor across macOS

What it does

Cotabby lives in your menu bar and drops AI-generated next-word suggestions as translucent ghost text wherever you type on macOS—Safari, Notes, Mail, no exceptions. Press Tab to accept, or keep typing to ignore it. It also handles :emoji: shortcuts inline, so you can insert 🎉 without switching apps.

The interesting bit

Rather than wrangling chat-tuned models, the open-source engine runs small base models—pure text continuers—through an in-process llama.cpp wrapper called CotabbyInference. It conditions them on your writing style, language, and even an OCR’d screenshot of the surrounding screen to guess what comes next. On macOS 26, you can opt into Apple’s FoundationModels runtime instead.

Key highlights

• Runs fully offline with downloadable GGUF models (Qwen3.5 and Gemma E2B variants, roughly 0.8 GB to 5 GB) or via Apple Intelligence on newer macOS
• System-wide ghost-text overlay works in any native text field without app-specific plugins
• Bring-your-own-model support: drop any small GGUF into the models folder
• Screenshot context gives the model visual awareness of what is on screen
• Licensed under AGPL v3

Caveats

• Requires broad macOS permissions: Accessibility, Input Monitoring, and Screen Recording access to read text fields, detect Tab presses, and capture surrounding context
• Apple Intelligence suggestions require macOS 26; earlier versions rely solely on local GGUF models
• The project is explicitly labeled beta

Verdict

Worth a look if you want Copilot-style suggestions in iMessage or Notes without subscription fees or data leaving your Mac. Windows and Linux users need not apply.