Get inline ghost text word predictions while typing in chat apps, writing apps, and browsers on macOS
Speed up repetitive typing like sign offs, names, and jargon using learned phrase memory
Run a private, local alternative to Cotypist without sending any keystrokes to the cloud
| nilava/tabtype | alamofire/rdar-urlsessiontaskmetrics | anywhere-music-player/codexlimits | |
|---|---|---|---|
| Stars | 5 | 5 | 4 |
| Language | Swift | Swift | Swift |
| Last pushed | — | 2017-12-15 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 2/5 | 1/5 | 2/5 |
| Audience | general | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires macOS 14 or newer on Apple Silicon, the app is unnotarized alpha software so macOS Gatekeeper may warn on first launch.
TabType is a free, open source autocomplete tool for macOS that predicts what you are about to type and shows the prediction as dimmed ghost text right where you are typing, in almost any application. You press Tab to accept a word at a time, or accept the whole suggestion at once. Everything runs locally on your Mac using a small open source language model called Qwen3, so no keystrokes are sent to the cloud, no account is required, and there is no subscription. The project describes itself as an open source alternative to a paid app called Cotypist. The suggestions try to fit the context of what you are actually doing. In chat apps like Slack, WhatsApp, Telegram, and Discord, and in chat websites like Claude and ChatGPT, TabType reads the visible conversation using the operating system's accessibility features rather than screenshots, though it falls back to reading the screen when that is not available. In writing apps like Pages, Word, and Notes, it reads a window of text around your cursor plus the start of the document so long documents stay on topic. A settings screen shows exactly what context each app is allowed to see, and you can change that behavior per app. The tool also remembers recent messages and phrases you type often, like names or sign offs, so its suggestions match your own writing style over time. Privacy is a core focus: inference happens entirely on your device, the only network call is a one time download of the language model from Hugging Face, optional typing history is encrypted locally, and password fields are never read. You can choose from a curated set of small local models or load your own from Hugging Face, and on newer Macs there is an option to use Apple's own on-device Apple Intelligence engine instead. Extra features include inline shortcuts like typing a date or converting units, emoji shortcuts, and autocorrect support for several Indian languages in addition to English. The project is an early alpha built mostly in spare time by one developer with heavy use of AI coding assistance, so it is described as useful but rough around the edges, and the author is actively looking for bug reports and contributions. It is released under the MIT license.
A free, open source, on device AI autocomplete for macOS that shows ghost text word predictions in almost any app, without cloud or subscription.
Mainly Swift. The stack also includes Swift, MLX, Qwen3.
Free to use, modify, and share for any purpose under the MIT license.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly general.
This repo across BitVibe Labs
Verify against the repo before relying on details.