What is Vox for Claude Code?

Vox is a local-first text-to-speech CLI for Claude Code. It reads task completions aloud, provides spoken summaries, and supports voice cloning from a 10-second audio sample. Three TTS backends are available: macOS say (~50ms latency), MLX-powered qwen neural (~500ms, Apple Silicon), and pure-Rust qwen-native (~300ms, Metal and CUDA GPU). Zero API calls, fully local, MIT licensed.

Does Vox send audio data to the cloud?

No. Vox is 100% local. All TTS processing runs on-device using local models. No API keys required, no data leaves your machine. The qwen-native backend is a pure Rust implementation with Metal and CUDA GPU support that works cross-platform without any Python dependencies.

How do I install Vox and connect it to Claude Code?

Run the one-liner installer: curl -fsSL https://raw.githubusercontent.com/rtk-ai/vox/main/install.sh | sh — then run vox init mcp to register Vox as an MCP server in Claude Code. This exposes 8 MCP tools including vox_speak, vox_clone_add, and vox_config_set.

What TTS backends does Vox support?

Vox supports three backends: (1) say — macOS native NSSpeechSynthesizer, ~50ms latency, instant startup, dozens of built-in voices; (2) qwen — MLX-powered neural TTS on Apple Silicon, ~500ms, high-quality speech with voice cloning, ~1.2GB model auto-download; (3) qwen-native — pure Rust implementation, ~300ms, cross-platform with Metal & CUDA GPU support, no Python dependencies.

How does voice cloning work in Vox?

Record a 10-second audio sample with vox clone record myvoice --duration 10, or register from an existing audio file with vox clone add --audio ~/voice.wav. Then use any cloned voice with vox -v myvoice "text". Voice cloning requires the qwen backend on Apple Silicon.

Open Source · MIT License · Written in Rust

Give Claude Code
a voice.

Hear when tasks finish, get spoken summaries, clone your own voice. Local-first TTS with 3 backends, zero API calls.

3 TTS backends

8 MCP tools

0 API calls

Vox runs entirely on-device — no API keys, no data leaving your machine. The say backend uses macOS NSSpeechSynthesizer for ~50ms latency with dozens of built-in voices. The qwen neural backend on Apple Silicon delivers natural-sounding speech at ~500ms via MLX, with voice cloning from a 10-second audio sample. The qwen-native pure-Rust backend runs cross-platform at ~300ms with Metal and CUDA GPU acceleration, no Python required. A single vox init mcp command registers all 8 MCP tools in Claude Code — vox_speak, vox_clone_add, vox_config_set, and more — enabling Claude to speak, manage voices, and report usage without any manual configuration.

Install Vox Star on GitHub

Why Vox?

Your AI assistant works in the terminal. Now it can talk to you.

Task Notifications

Hear when builds finish, tests pass or fail, and deployments complete. Stay productive without watching the terminal.

Spoken Summaries

Claude reads back what it did after a complex task. Get the gist without scanning hundreds of lines.

Voice Cloning

Record a 10-second sample and Vox speaks with your voice. Or clone any voice from an audio file.

100% Local

No API keys, no cloud, no data leaving your machine. Everything runs on-device with local models.

Cross-Platform

macOS, Linux, WSL. The pure-Rust backend works everywhere. Metal & CUDA GPU acceleration available.

One-Command Setup

vox init mcp registers everything in Claude Code. Zero config, instant integration.

3 TTS backends, one CLI

Pick the right engine for your platform and quality needs.

say macOS native

System speech synthesis via macOS NSSpeechSynthesizer. Instant startup, zero downloads, dozens of built-in voices.

Latency ~50ms

Quality System

Platform macOS

qwen Neural

MLX-powered neural TTS on Apple Silicon. High-quality, natural-sounding speech with voice cloning support. Model auto-downloads (~1.2 GB).

Latency ~500ms

Quality Neural

Platform macOS (Apple Si.)

qwen-native Pure Rust

Cross-platform neural TTS. Pure Rust implementation with Metal & CUDA GPU support. No Python, no dependencies.

Latency ~300ms

Quality Neural

Platform All

See it in action

Simple CLI, powerful results.

Basic usage

$ vox "Build complete. 3 tests passed."
# speaks aloud

$ vox -b qwen-native "Cross-platform TTS."
# uses pure-Rust backend

$ echo "Pipeline done" | vox
# reads from stdin

Voice cloning

$ vox clone record myvoice --duration 10
# records 10s of your voice

$ vox clone add patrick \
    --audio ~/voice.wav \
    --text "Transcription helps quality"
# register from audio file

$ vox -v patrick "Speaking with your voice."
# use cloned voice

Configuration

$ vox config set backend qwen
# switch default backend

$ vox config set lang fr
# set default language

$ vox config set voice Chelsie
# set default voice

$ vox --list-voices
# show available voices

Claude Code integration

4 init modes. Pick what fits your workflow.

mcp Default

Registers vox serve as an MCP server in ~/.claude.json. Exposes 8 tools that Claude can call directly.

vox init mcp

cli

Creates a CLAUDE.md with instructions and adds a stop hook that speaks "Terminé" when tasks complete.

vox init cli

skill

Generates a /speak slash command in ~/.claude/commands/. Invoke with /speak Hello.

vox init skill

all

Runs all three modes at once. MCP server + CLAUDE.md + slash command. The full experience.

vox init all

8 MCP tools exposed

vox_speakRead text aloud

vox_list_voicesList available voices

vox_clone_addAdd a voice clone

vox_clone_listList voice clones

vox_clone_removeRemove a clone

vox_config_showShow preferences

vox_config_setSet a preference

vox_statsUsage statistics

Install in seconds

Rust-powered. No runtime dependencies for the native backend.

Quick Install

macOS, Linux & WSL

curl -fsSL https://raw.githubusercontent.com/rtk-ai/vox/main/install.sh | sh

Via Cargo

From source

cargo install --git https://github.com/rtk-ai/vox

With GPU

Metal or CUDA acceleration

cargo install --git https://github.com/rtk-ai/vox --features metal

Then connect to Claude Code