Voice Box is an open-source voice-to-text tool that captures speech, transcribes it via Whisper, and formats output with an LLM. Press Ctrl+Cmd, speak, release — formatted text lands in your clipboard and auto-pastes into the active app. The architecture uses a Wails desktop app (Go + React WebView) streaming PCM audio over WebSocket to a Cloudflare Worker (Durable Object) running Whisper STT (@cf/openai/whisper-large-v3-turbo) and an LLM formatter (@cf/qwen/qwen3-30b-a3b-fp8).

The desktop client handles global hotkeys, audio capture (16k Hz mono PCM, ~4096-byte chunks), macOS accessibility for auto-paste, and a floating overlay UI showing voice level meter and processing state. Configuration lives in ~/.config/voicebox/voicebox.toml with cloud worker URL and shared secret. Audio streams up to ~25 MiB (~13 minutes) per session.

The Cloudflare Worker receives binary PCM chunks, accumulates audio, runs STT and formatting via Workers AI, and returns formatted text over WebSocket. The configure message includes focused element context (app name, bundle ID, element role) so the LLM tailors output. Local backends (faster-whisper, Ollama) are planned for Phase 2.

Development uses `wails dev` for hot reload, `wails build` for production, and `wrangler` for worker deployment. Requires Go 1.24+, Node.js, pnpm, Wails v2 CLI, and a Cloudflare account with Workers AI access. macOS accessibility permission is required for auto-paste.