Repository navigation
Releases
Bolkit 1.3.5 for Mac
Bolkit 1.3.5 for Mac
Dictation
- Take it back by voice. Say "no no no, delete everything I said" and only what you say after that is typed. "Start over" works too.
- Smart corrections. With an on-device AI model installed, slips like "let's meet at 2, actually 3" come out as "let's meet at 3". Bolkit only accepts the change when words were removed, never added or reworded. Turn it off in Text AI.
- Apple's Hinglish recognizer is now tried first in Auto.
- Your stats on Home. Words today, this week and all time, dictations, speaking speed, day streak, time dictated and meetings. A Cloud usage card counts your Azure requests, audio minutes and tokens. Only counts are kept, on this Mac.
Screen recording
- A picker you can trust. Screens and windows show live previews, grouped into Entire screen and Window. Point at one and the real screen or window lights up.
- You always know what is being recorded. A red border and a "Bolkit is recording" label sit on the screen or window being captured. They do not show up in the recording.
- Recorder window and call detection for meetings and screen recordings.
- Movable recording island that snaps to the notch, a corner or the bottom of the screen.
Meetings
- Make a video agent-ready. Import a Loom or screen recording as a meeting and get the transcript, speakers and summary.
- agent.md for every meeting, a timestamped transcript with screenshots from its own screen clips.
- Redo a meeting with a known number of speakers.
- Optional Azure cloud speech with your own key (MAI-Transcribe-2, real time, GPT-4o transcribe), with a backup resource when the first one fails.
Fixes
- Every second Apple dictation could come back empty.
- Parakeet runs on the CPU, so dictation is no longer stopped at the memory limit.
- A stop button could stay in the menu bar after a screen recording.
- Meetings over 2.5 hours could never finish saving.
- The island stays at the notch in full screen, and meetings can be deleted.
Requirements
- macOS 14 or later. Meeting recording needs macOS 14.4, screen recording macOS 15.
Bolkit 1.3.2 for Mac
Bolkit 1.3.2 for Mac
Meetings
- Hindi and Hinglish calls transcribe properly. Meetings used an English-only model by default, so Hinglish came out as made-up English words. Meetings now use Gemma 4 when it is installed, and the transcript is written in English letters. Apple speech is still there as an English-only option.
- In-person meetings are split by speaker. When nothing is playing on the Mac, everyone talking into your mic is separated into speakers instead of all being labelled You.
- Full-screen recorder. A dedicated recording screen with a big timer and live levels, from the Meetings card or the menu bar.
Fixes
- Starting a meeting recording could crash the app when the mic changed or reported a stale format. The mic now starts safely and restarts if you switch devices mid-meeting.
Bolkit 1.3.1 for Mac
Bolkit 1.3.1 for Mac
Scribe is now Bolkit, the same name as the iPhone app. Your settings, models, recordings and permissions carry over.
Meetings
- Apple speech for meetings. On macOS 26 and later, meetings are transcribed with Apple's on-device speech model, which hears each side of the call in one pass. It is the most accurate option we tested on English calls. Turn it off in Meetings to use your dictation model for Hindi or Hinglish calls.
- Better speaker separation. Up to 4 people on a call are separated automatically with NVIDIA Sortformer (downloads once, 242 MB). No need to pick a number first.
- No more echo without headphones. When your mic picks up the call from your speakers, Bolkit removes it before transcribing, so other people's words no longer show up twice under You.
- Meeting page. Open any meeting in the app: play it back, read the summary, search the transcript and rename speakers.
- Runs in the background. Meetings are transcribed in a queue with their progress shown on the meeting, and you can keep dictating or record another call meanwhile.
- Apple Intelligence summaries. When Apple Intelligence is ready, it writes the names and summary. Otherwise your AI Cleanup model does.
Fixed
- Gemma and the other on-device AI models now run on the GPU with MLX. In 1.3.0 using them could close the app.
- Long meetings no longer use up the Mac's memory (it could reach 13 GB).
- The app no longer closed while naming speakers on long meetings.
- Transcripts no longer include the model repeating its own instructions.
- Speakers are only named when someone actually introduces themselves or is named by someone else.
- Only one copy of each AI model is kept on disk.
Requirements
- macOS 14 or later. Apple speech and Apple Intelligence need macOS 26 or later.
Scribe 1.3.0
Scribe 1.3.0 for Mac
New
- Record meetings. Record any call from the menu bar or the dashboard. Your mic is saved as You and everything playing on the Mac (Zoom, Meet, Teams) as Others. After you stop, Scribe transcribes it on your Mac, separates the other speakers, puts real names on them when someone says their name, and writes a summary with key points, decisions and action items. You also get a playable
meeting.m4a. - Notch timer. While a meeting records, a small pill around the notch shows the timer and live waveforms for You and Others.
- New dashboard. Tabs for Home, Dictation, Meetings, Models, AI Cleanup, Files, History and Settings, with plain-language settings.
- Apple Intelligence. On macOS 26 and later, meeting names and summaries use Apple's on-device model when it is ready. Otherwise Scribe uses your AI Cleanup model.
- Automatic updates from this version on.
Changed
- The dictation pill now shows only a waveform and "Listening" while you talk. Text appears once, when it is inserted.
Fixed
- Audio models (Gemma 4 E2B Audio) sometimes wrote Scribe's built-in word list into transcripts.
Requirements
- macOS 14 or later. Meeting recording needs macOS 14.4 or later.
- The first meeting recording asks for System Audio Recording permission.
1.2.0 did not include the updater, so install 1.3.0 once by hand. Later versions update themselves.
Scribe macOS update feed
Stable Sparkle update feed. Managed by release automation.
Scribe 1.2.0
macOS
- Hinglish that reads right: new Srota model (Qwen3-ASR 0.6B fine-tune, on-device GGUF) plus a proper Devanagari-to-Roman transliterator with real schwa deletion, so Hindi comes out as "ab isko badalne ke liye" instead of Devanagari or gibberish. Roman output is on by default, toggleable.
- Spoken commands: "next line", "new paragraph", "point one / two / three" for numbered lists, "bullet" for dashes, "scratch that" / "cancel the last line". Deterministic and instant, no model involved.
- Instant capture: offline models open the mic immediately and load in the background; the recording starts the moment you hit the hotkey.
- Bluetooth mic fixes: the mic is fully released after each use (no more degraded headphone audio), capture survives device switches, and a watchdog restarts a silently muted input.
- Lower memory: every engine (ASR, Srota, Gemma cleanup) unloads after 5 minutes idle.
- Simpler model list: name, size, and a quality tag. Language picker and GPU toggle in settings.
Unzip, move Scribe.app to Applications, grant Mic + Accessibility.
Android APK and Windows build: see v1.0.0-beta for now; updated builds coming.
Srota Hinglish GGUF (Qwen3-ASR fine-tune)
GGUF conversion of moorlee/qwen3-asr-0.6b-hinglish (Apache-2.0) for Scribe's on-device Hinglish dictation. LLM q8_0 + mmproj audio encoder f16.
Scribe 1.0.0 (beta)
Scribe: 100% on-device speech-to-text. Your voice is transcribed locally and never leaves your device. This is an early beta: the binaries build cleanly but are not yet fully device-tested, and the desktop apps are unsigned (you'll need to bypass the OS security prompt once). Feedback welcome.
📱 Android: Scribe-android-arm64.apk (arm64 phones)
- Copy to your phone and tap it; allow "install from unknown sources" if asked.
- Open Scribe → grant Microphone.
- Voice keyboard: Settings → Open keyboard settings → enable Scribe, then tap 🌐 in any app.
- Flow Bubble (float a mic over any app): Settings → Flow Bubble → allow "display over other apps" + enable Accessibility. The bubble appears when you tap into a text field.
🍎 macOS: Scribe-macOS.zip
- Unzip → drag
Scribe.appto Applications. - It's unsigned, so first launch: right-click → Open (or run
xattr -cr /Applications/Scribe.app). - Grant Microphone, Speech Recognition, Accessibility when prompted.
- Hold the Right ⌥ Option key to dictate (configurable in Dashboard).
🪟 Windows: Scribe-windows-x64.exe (Windows 10/11, x64)
- Run it. SmartScreen will warn (unsigned) → More info → Run anyway.
- Look for the tray icon; use the global hotkey to dictate.
⚠️ Cross-compiled and not yet run on real Windows; please report whether it launches.
Known limitations
- Not device-tested end-to-end; desktop builds unsigned (no notarization / code-signing yet).
- Android APK is arm64-only (most modern phones). No iOS build yet.