v0.3.0 · streaming TTS, live carousel

Vox

Voice orchestrator for your coding agents.

A floating pill that watches your Conductor worktrees, briefs you out loud, and drafts the next prompt — hands-free, entirely on your Mac.

Download for macOS View on GitHub

macOS 12+ · Apple Silicon · MIT

⌥ space

Option + Space to talk. Vox listens, thinks, and answers out loud.

How it feels

You talk. It orchestrates.

FR« Trois worktrees actifs, un agent bloqué sur les tests d'auth depuis dix minutes — je le relance ? »

ENYou say what needs doing. Vox drafts the detailed prompt from the conversation and launches Claude Code on the right worktree.

FR/ENNo robotic roll call. Quiet worktrees get grouped, phrasing varies every time — « et sinon, rien de neuf côté billing » — like a teammate, not a cron job.

What it does

Built to sit next to Conductor, not replace it

Vox reads the state of your worktrees and gives you a voice interface to act on it.

Full worktree recap on launch

Reads Conductor's SQLite DB read-only and speaks a recap of every active worktree: who's working, who's idle or errored, and what each agent last said.

Voice-drafted prompts

Describe what needs doing. Vox turns the conversation into a detailed prompt and launches a Claude Code agent on the worktree you mean.

Streaming replies

New in v0.3.0: Vox starts speaking while the model is still generating, so a recap or an answer never feels like it's loading.

Real progress advice

Vox compares what an agent shipped against your original ask, tells you if it's on track, and proposes the next prompt to send.

Bilingual, end to end

Français and English throughout — STT, brain, and TTS. A pronunciation dictionary at ~/.vox/pronunciations.json fixes the words it mangles.

100% local

Whisper, Ollama, Kokoro/Piper — everything runs on your Mac. No API keys, no cloud calls, nothing leaves your machine.

Private by construction

Nothing you say leaves your Mac.

No accounts, no telemetry, no API keys to paste in. Vox runs a small local stack and talks to Ollama on localhost — that's the entire network surface.

Speech-to-textWhisper (local Python daemon)

ReasoningOllama — any model, default qwen2.5:3b

Voice — EnglishKokoro TTS

Voice — FrenchPiper TTS

Fallback voicemacOS say

Worktree dataConductor's SQLite DB, read-only

Get started

Three steps, all local

Download & open

Grab the .dmg from GitHub Releases. It's ad-hoc signed, not notarized — the first launch, right-click the app and choose Open to get past Gatekeeper.

First launch

Vox detects what's missing and auto-installs the local speech stack (Whisper, Kokoro, Piper) into ~/.vox/venv — a panel shows the progress. Just Python 3.11+ and an internet connection.

Config lives in ~/.vox/settings.json and ~/.vox/projects.json.

Install Ollama, pull a model

Install Ollama and run ollama pull qwen2.5:3b (or any model you prefer). Launch Vox, hit ⌥ Space, and start talking.

Stop context-switching between terminals.

One pill, every worktree, your voice.

Download Vox