Skip to content

Releases: nilava/TabType

TabType v0.1.1 — faster, smarter, stabler

Choose a tag to compare

@nilava nilava released this 18 Jul 18:08

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

⚠️ Still an alpha — bug reports are gold.

TabType v0.1.0 — first public alpha

Pre-release

Choose a tag to compare

@nilava nilava released this 16 Jul 10:53

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

  1. Download TabType-0.1.0.dmg below and drag TabType to Applications.
  2. 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
  3. Grant Accessibility permission (required) when prompted. Screen Recording is optional but recommended for better context.
  4. 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