Hushtype
Hold a key, speak, let go, and polished text appears in whatever app has focus. An open,
privacy-first alternative to Wispr Flow, built as a native macOS app in Swift and SwiftUI.
What it does
- Push-to-talk on Right Option (default), Right Command, or Fn/🌐, captured system-wide with a
CGEventTap.
Pressing any other key while holding cancels, so normal shortcuts keep working.
- Hands-free: double-tap the key and talk as long as you like; tap again to insert, Esc to cancel.
- Speech to text on this Mac with Apple's
SpeechAnalyzer. Audio is transcribed live while you hold the key,
so the transcript is ready almost as soon as you let go. Free, private, and works offline once the
voice model for your language has downloaded (setup does this for you).
- Cleanup with Claude Haiku, streamed: removes fillers, fixes punctuation, applies self-corrections
("at 5, no, 6 pm" becomes "at 6 pm"). Short phrases with no fillers or corrections skip Claude entirely, the
connection is opened while you're still talking, and if Claude hasn't started answering within 2 seconds the
raw transcript is inserted instead.
- Context awareness: the app you're in and the text around the cursor (read through Accessibility, never
from password fields) go to Claude, so it matches tone (casual in Slack, careful in Mail, literal in editors),
continues your sentence, and spells names already on screen. Those names also help the transcriber.
- Command mode: select text, hold the key, press Control too, and say "make this shorter" or "translate to
Hindi". The result replaces the selection. With nothing selected, Claude writes what you ask for.
- Personal dictionary fed to both the speech engine and Claude, and it learns from your corrections: fix
the spelling of a name right after dictating it and Hushtype remembers it.
- Snippets: say "my calendar link" and the full text is typed.
- History saved on your Mac, searchable in Settings, and ⌃⌘V pastes the last dictation again.
- Speed readout: Settings › Claude shows typical and 90th-percentile time from letting go to text.
- Auto-updates with Sparkle in signed releases (see RELEASING.md).
Build and run
Requires macOS 26+ and Xcode 26 (or the Swift 6.2+ toolchain). The only key you need is an Anthropic API key.
scripts/build-app.sh # builds dist/Hushtype.app
open dist/Hushtype.app
You can also open Package.swift in Xcode to edit and debug, but permissions such as the microphone need the
bundled app from the script, since a bare executable has no Info.plist.
On first launch a short guided setup walks through Microphone and Accessibility (each
step notices the grant by itself and moves on, and restarts Hushtype for you when macOS requires it), checks your
downloads the on-device voice model, checks your Anthropic key as you paste it, and ends with a practice field to try your first dictation.
Run it again any time from Settings › General.
macOS ties Accessibility grants to the code signature, and the default ad-hoc signature changes
with every build. To keep grants across rebuilds, create a self-signed code-signing certificate in Keychain Access
and build with HUSHTYPE_SIGN_IDENTITY="Your Cert Name" scripts/build-app.sh.
Layout
Sources/HushtypeCore/ pipeline logic: on-device speech-to-text, Claude cleanup (unit tested)
Sources/Hushtype/ the menu bar app: hotkey, recorder, overlay, paste, settings, Keychain
Tests/ HushtypeCore tests (swift test)
Resources/Info.plist app bundle metadata (LSUIElement, mic usage string)
scripts/build-app.sh builds, icons and signs dist/Hushtype.app
scripts/release.sh signed, notarized release with DMG and Sparkle appcast (RELEASING.md)
docs/ website, served by GitHub Pages at hushtype.chiranjeeb.dev
assets/social-preview.* social preview image (node scripts/render-social-preview.mjs regenerates the PNG)
Contributing
Issues and pull requests are welcome. See CONTRIBUTING.md for how to build, test and
send a change.
License
Hushtype is open source under the MIT License. Copyright (c) 2026 Chiranjeeb Jena.
