Fast, private, on-device dictation for Apple silicon Macs.

LocalFlow turns speech into polished text at the active cursor without sending audio or transcripts to a cloud service. Hold ⌥ Option + Space, speak, and release. Apple SpeechTranscriber handles transcription and Apple Foundation Models can clean the result before LocalFlow pastes it into the focused app.

What it does

• Streams speech recognition locally with Apple SpeechTranscriber.
• Uses Apple Foundation Models for optional two-second smart cleanup.
• Falls back to the raw transcript if cleanup misses its deadline.
• Pastes safely into the currently focused text field.
• Shows listening, processing, success, and failure states in a compact notch HUD.
• Supports click-to-start/stop, push-to-talk, Reduce Motion, and a persistent 220–420 pt HUD width.
• Never stores audio, transcripts, prompts, or clipboard contents.

Requirements

• Apple silicon Mac (MacBook Air, MacBook Pro, Mac mini, Mac Studio, or iMac)
• macOS 26.0 or newer
• Xcode 26 or the matching Swift 6.2 command-line tools
• Microphone, Speech Recognition, and Accessibility permissions

Quick start

git clone https://github.com/codejunkie99/LocalFlow.git
cd LocalFlow
./scripts/package-app.sh
open dist/LocalFlow.app

The packaging script automatically uses the first installed Apple Development certificate. If none exists, it creates an ad-hoc signed local build and explains the permission trade-off.

For a full first-run walkthrough, see Installation.

Use

1. Focus a text field in Notes, TextEdit, Safari, Slack, or your editor.
2. Hold ⌥ Option + Space.
3. Speak naturally.
4. Release the keys to transcribe, optionally clean up, and paste.

A quick tap of the shortcut enables toggle mode: tap once to start and again to stop. You can also use Start Dictation in the LocalFlow window.

Privacy model

• No networking client is linked or used by the application.
• Audio exists only in memory while recording.
• Transcripts and cleanup prompts are never persisted.
• Logs contain state and timing information only.
• Clipboard contents are restored after the paste attempt when possible.
• Apple may download system-managed speech and language-model assets.

See Architecture for the trust boundaries and data flow.

Build and test

swift build --disable-sandbox
./scripts/run-tests.sh

Create a release application bundle:

./scripts/package-app.sh

Use a specific signing identity when needed:

LOCALFLOW_SIGNING_IDENTITY="Apple Development: Your Name (TEAMID)" ./scripts/package-app.sh

Stable signing lets macOS keep the Accessibility grant across rebuilds. An ad-hoc signature is fine for a quick local build, but macOS may ask for Accessibility again after the executable changes.

Performance target

For 20 warm trials:

• Median total latency: ≤ 1,500 ms
• p95 total latency: ≤ 2,500 ms
• Paste latency: ≤ 100 ms

Use artifacts/performance/acceptance-template.json with scripts/acceptance-stats.sh to record a local run.

Project layout

Sources/LocalFlowApp       SwiftUI menu-bar app and notch HUD
Sources/LocalFlowCore      State machine, deadlines, and privacy-safe policies
Sources/LocalFlowPlatform  Apple Speech, Foundation Models, shortcut, and paste adapters
Tests                      Deterministic core and platform checks
scripts                    Build, package, test, and acceptance helpers

Contributing

Bug reports and focused pull requests are welcome. Read CONTRIBUTING.md before making changes. Security and privacy issues should follow SECURITY.md.

License

LocalFlow is available under the MIT License.