Troubleshooting & FAQ
Open Diagnostics from the menu-bar icon — Context shows what Pretype sees (app, window, field) and Pipeline shows what the last completion did.
Common issues
Accessibility: NOT granted ✗ — if you’re running the raw binary from a terminal, macOS attributes the permission to the terminal, so grant it there or run the .app bundle. If you built locally, a changed code signature can confuse macOS:
tccutil reset Accessibility app.pretype.Pretypethen re-grant.
Text element: none — the app doesn’t expose its text fields over the Accessibility API. Nothing to be done from this side.
Last: engine returned no suggestion — the model had nothing it was willing to guess: too little context, or the output gates rejected what it produced. Keep typing.
No suggestions anywhere, MLX engine missing — a source build without compiled Metal shaders. Use ./Scripts/make-app.sh rather than plain swift build.
Gatekeeper blocks the first launch — releases before the Developer ID signing are ad-hoc signed. Clear the quarantine flag with xattr -dr com.apple.quarantine /Applications/Pretype.app, or open the app via System Settings → Privacy & Security → Open Anyway.
FAQ
What happens to Tab when I actually want a Tab?
It passes straight through. The event tap only swallows the key while a suggestion is on screen; with nothing showing, Tab indents and moves between fields as usual. If that still collides with your habits, switch the chord to ⌘Space, ⌥Space or ⌃Space in Settings → General.
Suggestions stopped appearing in one app
Deliberate. Once Pretype has offered a few dozen completions in an app and under 2% of them were taken, it stops offering there — a model that is wrong in a particular field is pure interruption, and generating them costs battery. Everything else keeps working: inline typo fixes, emoji shortcodes and ⌥Tab rewrites are unaffected.
Only offers you could actually have taken count toward that. The pipeline re-offers after every keystroke, so a ghost that was replaced by the next one, or that vanished in under 0.4 s, was never a suggestion you declined — and one you typed out yourself, word for word, means the model was right. None of those reach the tally, and neither do they reach the acceptance figure in Diagnostics.
It isn’t silent about it. The menu-bar item that normally offers Disable in … becomes Resume in Slack instead, hovering the numbers behind the decision (“Pretype went quiet here: 2% of 74 suggestions taken.”), and one click clears the record and starts offering again from scratch.
What does it cost in battery and memory?
The model stays resident while you’re typing — that’s what makes warm completions fast — holding ≈1.6–2.2 GB for the defaults. After five idle minutes it unloads itself and frees that memory, reloading on your next keystroke; it also unloads early if macOS reports memory pressure. It only computes while you’re typing in a field it’s allowed to read, and suggestions are debounced and cancellable, so fast typing supersedes in-flight work instead of queueing it.
Which languages work?
Completions are evaluated across 19 languages, each measured on its own rather than averaged with its neighbours — see Choosing a model; the default is picked from your keyboard layouts and the Gemma 4 builds have the broadest coverage. In Settings → Model you can set the accuracy axis to the language you actually type in, and every figure on that tab is re-measured for it, with its sample size and tolerance printed alongside. Inline typo fixes use the macOS spell-checker in whichever language it detects, so any dictionary macOS has installed works (English and Russian are the most-tested pair).
Where does my data go?
Nowhere. See Privacy & permissions for what is stored locally and how to remove it.
Still stuck? Open an issue with the Diagnostics output — it redacts your text and carries only the pipeline state.