Speak, capture, and craft prompts.
Vox is a voice-first workspace for developers who want to drive agents by prose. On-device Whisper transcription (no audio leaves your machine), meeting capture that records mic and system audio, searchable local history, versioned prompts with AI coaching, and a custom dictionary for deterministic corrections.
The problem
Developers work with AI agents every day, but the interface is text: you type a prompt, the agent responds. This creates cognitive friction — you lose the natural rhythm of speech, context gets trapped in local variables and inarticulated assumptions, and by the time you've typed a prompt, you've forgotten half of what you wanted to say. Recording a voice memo is faster, but the transcript needs cleanup. Cloud transcription means your code and context flow to external servers. You need transcription that stays local and is fast enough to feel immediate.
What Vox does
Vox brings voice-first interaction to the developer workspace. Whisper transcription runs entirely on your machine via whisper.cpp — you pick your model size (tiny through large, trading speed for accuracy) and nothing leaves your device. A searchable local SQLite history stores every transcript and prompt. Meeting Mode captures both your mic and your system audio (capturing what's being said on calls or in recordings) into chunks you can bookmark and reference. The Prompt Workspace lets you version and refine prompts over time, with optional AI coaching from OpenRouter or Anthropic to help you craft better instructions. A Custom Dictionary learns your terminology — tool names, API conventions, project jargon — so the transcriber stops fumbling over "Codex" or "async-await" and gets it right every time.
What makes it different
- On-device transcription. Whisper.cpp brings speech-to-text to your machine. You control which model to load (tiny through large). No audio ever touches an external server.
- Meeting Mode. Capture mic input plus system audio simultaneously — perfect for recording calls, walkthrough sessions, or talks you want to reference later. Store and search everything locally.
- Versioned prompts. The Prompt Workspace lets you evolve your prompt instructions over time. Each version is immutable; you can compare and branch. Optional AI coaching helps you refine your technique.
- Custom dictionary for deterministic corrections. Teach Vox your project's terminology — function names, tool abbreviations, domain terms — so corrections are predictable and reproducible.
- Disciplined architecture. 44 architecture decision records, a domain glossary, module/adapter/store vocabulary, and swappable backends. Built to be understood and extended.
Status
Vox is in the works. The transcription pipeline, meeting capture, and local history search are complete and tested. Prompt versioning and the custom dictionary are in active development. The first release is in preparation and will be available in Q4 2026.