Architecture

VoxFlow is a Rust core with a Tauri 2 + React shell on top. The split is deliberate: anything with a latency target lives in Rust, and the web layer only renders state that has already been decided.

The pipeline

Hotkey → Audio capture (CPAL) → VAD (WebRTC) → Whisper (local)
       → AI rewrite → Insert at cursor → History

Crates

Crate	Role
voxflow-core	Dictation state machine and pipeline orchestration
voxflow-audio	CPAL capture on a background thread
voxflow-vad	WebRTC voice activity detection, silence trimming
voxflow-whisper	Local transcription via whisper-rs with Metal
voxflow-provider	Generic OpenAI-compatible HTTP adapter for the rewrite pass
voxflow-router	Routing rules across configured providers
voxflow-insert	Text insertion trait and platform implementations
voxflow-history	SQLite transcripts and usage records
voxflow-cost	Cost tracking, caps, projections
voxflow-config	Settings JSON and per-app profiles
voxflow-secrets	Keyring-backed secret storage
voxflow-platform	Platform-specific glue
voxflowctl	Command-line utility, including key management

The app shell

• src-tauri/ — the Tauri host. Owns the hotkey listener, the tray icon, and two windows: a hidden settings window and a transparent always-on-top overlay. Exposes commands to the frontend through invoke and pushes state to it through events.
• src/ — React and TypeScript. The settings pages, and the overlay pill with its live waveform.

The overlay receives amplitude samples as events and draws them. It is never asked to make a decision, because a WebView cannot be trusted with an 80 millisecond budget.

Design rules

The product spec sets constraints the implementation is held to:

• Never steal focus unless it is unavoidable.
• Never open a large window during dictation.
• Never make the user paste manually unless insertion genuinely failed.
• Always give a fallback path.
• Always prioritise perceived speed.