Who
I build local tooling when cloud dictation is the wrong trust boundary.
For operators who want Ctrl+W anywhere to type into the focused field with whisper.cpp on the machine, and who need the honest split between the streaming HEAD and the batch/orange README still sitting in the same private repo.
What
I built WhisperCPPonEverything, private michaelmonetized/WhisperCPPonEverything. Native macOS menubar (LSUIElement) Swift Package executable. HEAD d33e5e2. 11 commits. 0 stars. Bundle com.whispercpponeverything.app 2.0.0. Renamed from VoiceType.
Stack facts from Package.swift + sources:
- Platforms: macOS 14+
- Frameworks only: AppKit, CoreGraphics, ApplicationServices. Zero SPM dependencies.
- Hotkey: Ctrl+W via
CGEventtap (active session, then listenOnly, then HID listenOnly fallback) - macOS 15+:
CGPreflightListenEventAccess/CGPreflightPostEventAccess+ request APIs before tap create - Transcription: subprocess
/opt/homebrew/bin/whisper-streamwith model/opt/homebrew/share/whisper-cpp/models/ggml-medium.bin, args-l en --step 3000 --length 5000 --keep 200 - Incremental inject: strip ANSI +
[timestamp]prefixes, suffix-prefix overlap dedupe, thenTextInjectorUnicodeCGEventkeystrokes (5ms between chars) - Visual: fullscreen
.screenSaveroverlay, cyan 20-layer soft glow, 1.5Hz pulse,ignoresMouseEvents - State machine:
idle/streamingonly (State.swift) - Logging:
~/Library/Logs/WhisperCPPonEverything.log - Checked-in
WhisperCPPonEverything.appwith arm64 Mach-O binary
What the README still claims (stale vs HEAD):
- Orange screen glow while listening
- Silence detection ends recording after 5 seconds
- Batch-style speak, stop, then transcribed text is typed
What CLAUDE.md + Swift say at HEAD: cyan soft glow, toggle stop, streaming text as whisper-stream prints lines. PLAN.md still names AudioRecorder.swift / Transcriber.swift. Those files are not in the tree. TODO.md phases are all unchecked. StatusItem streaming mic is still .systemOrange next to a cyan overlay.
Distinct from niri-macos (Swift AX tiling) and from SaaS dictation landers. Local Homebrew whisper-stream or the app alerts that whisper-cpp is missing.
Where
Code: github.com/michaelmonetized/WhisperCPPonEverything, private.
Runtime probes at pack time: no public web product. Install path documented as swift build -c release then copy into /Applications/WhisperCPPonEverything.app/Contents/MacOS/ + ad-hoc codesign. CLAUDE note: each new binary invalidates TCC. Remove and re-add Accessibility.
Required on disk: whisper-stream + ggml-medium.bin under Homebrew paths above.
When
- 2026-02-15. PLAN/README/TODO, rename, sources build, VoiceType rename, whisper-cli path fixes, app bundle, cyan instead of orange.
- 2026-02-16.
92bb411streaming STT + soft glow + macOS 15+ permission support. - 2026-06-22. Two
nightlycommits; HEADd33e5e2(also last GitHub push).
Why
I wanted dictation that stays on-box and types into whatever already has focus (Slack, terminal, browser) without a cloud STT round-trip. Ctrl+W was the whole UX. Streaming beat record-a-WAV-then-wait. Cyan beat ugly orange. Sequoia event-access APIs were the tax for keeping the tap alive.
Engagement Q: How many private STT menubar apps still advertise orange silence-batch in README while HEAD streams cyan into the focused field?
Draft + assets only until Michael publishes.