Repository navigation
Releases
TabType v2.1.0
TabType v2.1.0 closes most of the gap with Cotypist: ghost text that sits where the app's own text would, suggestions that read the conversation you're replying to, longer phrases, and keys that never get in your way.
Every change was measured: on the public 192-case eval set, and on a private set of 867 cases from the author's own messages. What we found and why each change was made is in the README: What changed in v2.1 — and why.
📈 Highlights
| Before | v2.1 | |
|---|---|---|
| Suggestions with more than one word (author's writing) | 22% | 57% |
| Characters saved per suggestion (author's writing) | 1.37 | 1.46 |
| Placements using the app's real font | 0 of 984 | read from the app |
| Chat predictions with conversation context (Slack, live log) | 72 of 406 | captured above the field |
| First suggestion after launch | ~2.2 s | ~0.24 s |
| Suggestions after the Mac sat idle | stopped | reload automatically |
We also tried Cotypist's own model (Gemma 4 E2B) inside TabType: about 50 ms faster, but 20% fewer characters saved on real writing, so Qwen3-4B stays the default.
✍️ Placement
- Ghost text uses the app's own font and colour (read from accessibility, with pixel fitting as a fallback), and wraps onto the next line like real text.
- Works in more editors: hidden-input web editors, visual line starts, apps that report the caret late (retry plan + line-height cache).
- Text mirror: a per-app opt-in preview for apps where inline ghost text can't be placed.
- The ghost appears as soon as the app has caught up with your typing, with no fixed delays.
🧠 Better suggestions
- Longer phrases: a beam search over whole phrases (like Cotypist's) and a lower extension bar. Suggestions average ~2 words instead of mostly one, with the first word unchanged.
- Conversation context: reads the chat or document above the field with on-device OCR (accessibility tree as a fallback), and keeps that context steady while you type a message.
- Typing ahead of a running suggestion reuses it instead of starting over; recent results are cached; instant Tab chains.
- Learn from your writing imports your existing history once when you turn it on, with a strength setting and an "only record messages where I used a suggestion" option.
- Typo gate on the word you're typing; small fields (search boxes) are skipped.
⌨️ Keys
- TabType only listens to the keyboard. Tab, ⇧Tab and Esc are taken only while a suggestion is showing; ⌥Tab sends a real Tab.
- A notice when another app already owns a shortcut.
✨ New
- Terminals: suggestions inside AI agents' prompts (Claude Code, Codex, Gemini CLI), never at the shell.
- In-app updates from GitHub Releases, verified by checksum and signature before installing.
- Sync between Macs via iCloud Drive, end-to-end encrypted with your passphrase.
- On battery power: on demand only, shorter suggestions, or a smaller model.
- Per-app insertion workarounds for apps that mangle inserted text; grey suggestion colour option.
- Labs: word alternatives and synonyms after a pause. Emoji: preferred gender and gender-neutral variants.
🐛 Fixes
- Suggestions no longer stop after the model is unloaded while idle or under memory pressure.
- The first suggestion after launch no longer waits ~2s for the spelling dictionary; the first keystroke in a new field predicts at once.
- The diagnostic log is capped at 5 MB, and can be deleted from Settings → Advanced.
- Interrupted model downloads resume at launch.
🧹 Removed
- Settings that got in the way of suggestions: suggestion delay, context window size, screenshot mode, the voice adapter. (On-battery options came back in a simpler form.)
Install / Update
Download TabType-2.1.0.dmg below and drag it to Applications, replacing the old copy. Your permissions are preserved. From this version on, TabType can update itself (Settings → About).
New install? macOS will warn about the unnotarized app: System Settings → Privacy & Security → "Open Anyway", then grant Accessibility and Screen Recording (recommended: it's how TabType reads the conversation you're replying to). Full steps are in the README.
Apple Silicon Mac, macOS 14+.
SHA-256 bf0a13cbb8d9e880f9af49447a9da30341a1645771fcbbe94b2406e8656ef90f
TabType v2.0.0 — new llama.cpp engine, confidence-gated suggestions
TabType v2.0.0 replaces the completion engine from the ground up. v1 asked a chat model to "continue the user's text", and too often it replied, rambled, or guessed. v2 runs a base language model through llama.cpp: it simply continues your writing, and a new decoder only shows a suggestion when the model is actually confident.
📈 Measured, not guessed
On TabType's offline eval set (192 real-world typing cases, tabtype-eval):
| v1 | v2 (Qwen3-4B base) | |
|---|---|---|
| Next word correct | 37% | 56–58% |
| Precision (shown suggestions that were right) | 38% | ~70% |
| Wrong suggestions shown | 61% | ~22% |
| Median latency | ~300 ms | ~80–100 ms |
🧠 New engine
- llama.cpp + GGUF base models. Qwen3-4B on 16 GB+ Macs, Qwen3-1.7B on 8 GB. Gemma 4 and smaller Qwen models are selectable in Settings → Model.
- Confidence-gated decoding. Every suggestion carries the model's own probability. Weak guesses are never shown, and multi-word phrases only extend while confidence holds.
- Token healing. Mid-word suggestions finish the word you're typing instead of starting a new one.
- Prefix-cached prompts. Context is laid out so the model reuses its cache across keystrokes, which keeps it fast.
✍️ Looks native
- Font-fitted ghost text. TabType renders candidate fonts and matches them against the field's pixels, so the ghost uses the app's own font, size, and baseline.
- Tab-accept keeps the rest of a suggestion in place instead of recomputing it.
🙋 Learns from you (opt-in)
- Learn from your writing. An encrypted, on-device index of what you've written nudges suggestions toward your names, phrases, and sign-offs. It's off by default and can be turned on per app. Erase it any time in Settings.
- Voice adapter. Advanced users can load their own LoRA adapter (GGUF) to steer style.
🧹 Removed
- MLX and the Apple Intelligence engine are gone. Base models complete text measurably better, and the app no longer has any third-party Swift dependencies.
- Temperature / max-token settings were removed; the decoder decides length from confidence.
- Phrase memory and the old typing history are replaced by Learn from your writing, which imports your existing history once.
Install / Update
Download TabType-2.0.0.dmg below and drag it to Applications, replacing the old copy. Your permissions are preserved.
Updating from 0.1.x: v2 uses a new model format, so it downloads a new model on first launch (~1.1–2.5 GB, depending on your Mac's RAM). To reclaim space, delete the old mlx-community… folders inside ~/Library/Application Support/TabType/Models (keep the .gguf files: that's v2's model) and any models--mlx-community… folders in ~/.cache/huggingface/hub.
New install? macOS will warn about the unnotarized app: System Settings → Privacy & Security → "Open Anyway", then grant Accessibility. Full steps are in the README.
Apple Silicon Mac, macOS 14+.
SHA-256 9509a310ce630535fccf24b5341a454c3fb7760b58afc5440867b7aa686d7bb5
TabType v0.1.2 — download & crash fixes, faster suggestions
TabType v0.1.2 — fixes two bugs that could stop TabType working entirely, and makes suggestions much faster and more relevant. Recommended for everyone, especially if the model never finished downloading for you.
🛠 Fixed
- Model download stuck at "Downloading 0%" forever. Two separate bugs: the progress handler was silently discarded, so the UI never moved; and the stall watchdog measured a directory that cannot grow while a multi-GB file is in flight, so it cancelled perfectly healthy downloads after 10 minutes — and since a cancel threw away the partial file, every retry restarted from zero. If your connection couldn't pull the whole model in one 10-minute run, TabType could never install a model at all.
- Downloads now resume. Files stream into TabType's own store and continue with an HTTP Range request, so a dropped connection, a quit, or a relaunch keeps every byte already transferred. Retries that made progress no longer count against the retry limit, so flaky connections eventually finish.
- Crash when the Mac's power state changed (Low Power Mode, plugging in/out) — a main-actor isolation trap in the power monitor.
- Suggestions replying instead of completing. With a conversation on screen the model would sometimes answer the other person rather than finish your sentence. Prompting is now strictly continuation-only, with guards at the sampler.
- Weights disappearing. Models were kept in
~/Library/Caches, which macOS purges under disk pressure. The store moved to Application Support; already-downloaded models are detected in either location, so nothing re-downloads on update.
⚡ Faster
- Suggestion latency in chat apps drops from ~0.8–1.5 s to ~160–270 ms. Conversation transcripts are normalized (timestamps, presence, "(edited)" stripped) and only re-read when the conversation actually changes, which keeps the model's cache warm between keystrokes — 99% prefix reuse.
- Screen captures pause while you type, and switching apps no longer causes a slow first suggestion.
🧠 Smarter context
- 15+ chat apps: adds Telegram, Signal, Teams, Messenger, Element, Beeper, Mattermost, Zulip, Skype and Google Chat to the apps whose conversation TabType reads via accessibility.
- Web chats: claude.ai, ChatGPT, Gemini, WhatsApp Web, Discord, Slack, Perplexity and more now work the same way inside any browser.
- Long-form writing: in Pages, Word, Notes, Ulysses, Bear and friends, TabType reads a much larger window around your cursor plus the document's opening lines, so suggestions deep in a document stay on topic.
- Context from another app — or another browser tab — is now labelled as background instead of being mistaken for what you're writing now.
🔍 Transparency
- Settings → Apps shows each app's profile (Chat / Document / Code editor) and explains in plain English exactly what TabType reads there — plus new per-app controls to turn conversation reading on for any app, or resize its context budget.
Install / Update
Download TabType-0.1.2.dmg below and drag it to Applications, replacing the old copy — your permissions and downloaded model are preserved. New install? macOS will warn about the unnotarized app: System Settings → Privacy & Security → "Open Anyway", then grant Accessibility. Full steps in the README.
SHA-256 ce7e729d1f6ef38081fc4ce9b87fe8998ff14c66a38d7383814f3c2638289700
TabType v0.1.1 — faster, smarter, stabler
TabType v0.1.1 — a big quality release: suggestions are 4–6× faster, understand far more context, and a crash is fixed. Recommended update for everyone on v0.1.0.
🛠 Fixed
- Crash when the Mac's power state changed (Low Power Mode toggle, charger events) — a main-actor isolation trap in the power monitor. If v0.1.0 quit on you mysteriously, this was it.
- Reply-drift: the model sometimes answered the conversation on screen instead of completing your sentence. Prompting is now strictly continuation-only, with sampler-level guards for reply openers.
- Suggestions no longer leak context between apps (or between browser tabs) — what you just read elsewhere is now explicitly labeled background, never mistaken for what you're writing now.
⚡ Faster
- Suggestion latency down from ~0.8–1.5 s to ~160–270 ms in chat apps: transcripts are normalized (timestamps/presence/"edited" junk stripped) so the model's KV cache stays warm between keystrokes, and screen captures pause while you type.
- Switching apps no longer causes a slow first suggestion — the new app's context is pre-warmed in the background.
🧠 Smarter context
- Reads the whole visible conversation in 15+ chat apps — now including Telegram, Signal, Teams, Messenger, Element, Beeper, Mattermost, Zulip, Skype, and Google Chat.
- Web chats too: claude.ai, ChatGPT, Gemini, WhatsApp Web, Discord, Slack, Perplexity and more get conversation-aware suggestions inside any browser.
- Long-form writing: in Pages, Word, Notes, Ulysses, Bear, and other writing apps, TabType reads a much larger window around your cursor plus the document's opening lines, so mid-document suggestions stay on topic.
- On 24 GB+ Macs, a higher-precision 8-bit model is now recommended automatically (same speed, better accuracy).
🔍 Transparency
- Settings → Apps now shows exactly what TabType does in every app: profile badges (Chat / Document / Code editor), a plain-English "How TabType works here" card, and per-app controls to change the context recipe — promote any app to conversation reading, or resize its context budget.
Install / Update
Download TabType-0.1.1.dmg below and drag to Applications (replace the old copy — your permissions are preserved). New install? macOS will warn about the unnotarized app: System Settings → Privacy & Security → "Open Anyway", then grant Accessibility. Full steps in the README.
SHA-256 95bd022e8a1160577217d8ac0ab9d2a23e0d936c67076d4540ca76a4cf3150c4
TabType v0.1.0 — first public alpha
TabType v0.1.0 — the first public alpha. A free, open-source, 100% on-device AI autocomplete for macOS (a Cotypist alternative).
Highlights
- Inline ghost-text completions everywhere you type — press Tab to accept word-by-word or all at once
- 100% local: a Qwen3-4B model runs on-device via MLX; no account, no cloud, no keystroke ever leaves your Mac
- Context-aware: reads the conversation around your cursor (accessibility tree in chat apps, on-screen text elsewhere), your recent messages, and learns your writing style
- RAM-tiered models: 24 GB+ Macs get a higher-precision 8-bit build automatically; a curated catalog (plus custom Hugging Face repos) is available in Settings
- Per-app control: chat-panels-only mode in code editors, per-app pause, per-domain website rules
- Typo-aware suggestions, word alternatives (⌃⌥Space), macros, emoji, autocorrect
Install
- Download
TabType-0.1.0.dmgbelow and drag TabType to Applications. - TabType is not notarized (no Apple Developer account — community help welcome!). On first launch macOS will warn you:
open System Settings → Privacy & Security and click "Open Anyway".
Power users can instead run:xattr -dr com.apple.quarantine /Applications/TabType.app - Grant Accessibility permission (required) when prompted. Screen Recording is optional but recommended for better context.
- First launch downloads the AI model (~2.3–4.3 GB from Hugging Face, one time). Suggestions start when the menu-bar icon shows ready.
Notes
⚠️ This is an alpha. Expect rough edges — bug reports are the most valuable contribution right now.- Full AI disclosure, feature comparison vs Cotypist/Sombra/KeyType, and architecture docs are in the README.
SHA-256 9e209d2a24f90355af26e29f2c01a7596566adca1405acb4939c138d7d42b983