Changelog
What changed in Kivo, what was fixed and what is still missing. Newest first.
Next
In development
Since 26 August 2026
These changes are finished but not in a release yet. The current download does not include them.
New
- A redesigned interface in black and white.
- The menu bar icon now animates while Kivo records and transcribes.
- Kivo warns you when a hotkey collides with a macOS default, and detects keys macOS has already claimed.
- Space is now available as part of a key combination, and Fn is named correctly.
Fixed
- The recording overlay could take focus from the app you were typing in.
- Settings could freeze and show “not responding”.
- A window could open blank.
- A crash while dictating could stop dictation until restart and leave the microphone open. The microphone now closes and Kivo carries on.
- Either Option key now works, not only the right one. The same goes for other modifier keys.
- Debug logs recorded the code of every key pressed while Kivo ran. That logging is gone.
- Model download now shows real progress, retries once and explains a failure plainly.
0.1.0
Latest
26 August 2026
Download for MacThe first build you can download and run. It is a preview for Macs with Apple silicon or Intel, on macOS 11 or later.
New
- Hold Right Option, speak and let go. Punctuated text appears at your cursor, in the app you are using.
- Speech recognition runs on your Mac using the Kivo v1 speech model.
- Fillers and stutters are removed, and single-word corrections are resolved. Each rule can be switched off, and Raw mode skips them all.
- Add your own names and jargon to the vocabulary.
- Double-tap to dictate hands-free. Press Esc to discard a recording.
- Searchable history, stored on your Mac. Turning it off deletes the file.
- The microphone opens on key-down and closes on key-up.
- The 460 MB speech model downloads once and is checked against a pinned SHA-256 hash.
Known limits
- Unsigned, so macOS blocks the first launch. Open it from System Settings, under Privacy & Security.
- Recognition took about 660 ms for 7.4 seconds of speech, against a 300 ms target.
- Uses about 960 MB of memory with the model loaded, against a 150 MB target.
- History is not encrypted. Anyone who can read your user account can read your transcripts.
- Text is inserted by pasting, and your clipboard is restored afterwards.
- English only and Mac only. Windows is planned, not started.