Changelog

What changed

Every release, line by line. Updates ship through Sparkle — one click inside the app.

Help us make Kept better

Plenty of the entries below started life as one of your emails.

v0.9.0

LatestDownload DMG ↓
today · July 25, 2026 · macOS 15+

The first public beta. Meetings record themselves, Chinese and English both transcribe on your Mac, and every line of the notes clicks back to the audio — none of it leaves the machine. Until 1.0 the cadence stays like this: fixes ship as they land, features as they finish.

New
  • Recording starts when you join and stops when everyone leaves, with a small floating bar and live captions while it runs. Waiting rooms never trigger it.
  • On-device transcription — FireRedASR2 for Chinese, Parakeet for English. Every word lines up with the audio, click one to jump there, and mixed Chinese-English stays intact.
  • It tells who said which line: 88.0% speaker accuracy on English, 82.2% on Chinese, measured on 25 hand-labelled real meetings (15.2 hours).
  • Enroll a colleague's voice once and their name shows up in every meeting after that. Below the confidence threshold it writes "Speaker 1" instead of guessing.
  • AI notes that write outcomes only: what was decided, who owns it, when it's due, what's still open. Click any line to land on the exact words.
  • A one-page visual summary you can copy as a single image and paste straight into an update.
  • Two-track lossless recording: FLAC, levels evened out, tracks aligned, and no double-recording the other side on speakerphone.
  • Three ways to record (mic only / mic + one app / mic + everything) and import of existing recordings (M4A, MP3, WAV, MP4, MOV, queued in batches).
  • Search titles, transcripts and notes together, then click a result to land on that moment in the audio.
  • Export to Markdown, TXT or a mixed m4a; copy the notes as text in one click.
  • Claude Code and Codex can read your meetings: one-click setup, read-only, text-only, off by default — switch it off and access stops.
  • It remembers where you stopped listening and resumes there, and the menu bar keeps recording after you close the window.
  • Speech models download themselves: resumable, checksum-verified, fastest mirror picked automatically. Once installed you can stay offline for good.

Earlier internal builds were never published · checksums on the download page