vox

A native macOS MCP server that gives Claude a voice.

Built in Swift. No Node.js. No Python. No Electron. Just a single binary.

{
  "mcpServers": {
    "vox": {
      "command": "/path/to/vox"
    }
  }
}

Then just say: listen — and Claude hears you.

Tools

Related MCP server: whisper-windows-mcp

Why Swift + macOS only

• AVFoundation — native mic capture, zero overhead

• SFSpeechRecognizer — Apple on-device speech recognition, works offline, 50+ languages

• NLLanguageRecognizer — automatic language detection per utterance

• AVSpeechSynthesizer — built-in TTS fallback, no API key needed

• Single ~200KB binary. Zero npm install. Zero Python venv.

Requirements

• macOS 13+

• Microphone set as default in System Settings → Sound → Input

• Microphone + Speech Recognition permission for your terminal app

• Optional: ELEVENLABS_API_KEY in ~/.claude/.env for high-quality multilingual TTS

Install

Option 1: Prompt for Claude Code

Copy and paste this into Claude Code:

Install vox (native macOS voice input/output MCP server):

1. Download the code-signed binary to ~/vox:
   curl -L https://github.com/boska/vox/releases/download/v1.1.0/vox-1.1.0-darwin-arm64 -o ~/vox && chmod +x ~/vox

2. Add to ~/.claude.json in the mcpServers section:
   {
     "mcpServers": {
       "vox": {
         "command": "/Users/$(whoami)/vox"
       }
     }
   }

3. Restart Claude Code

4. Test it by saying: listen

Claude will handle the download, configuration, and restart.

Option 2: Manual Install (Pre-built Binary)

curl -L https://github.com/boska/vox/releases/download/v1.1.0/vox-1.1.0-darwin-arm64 -o ~/vox && chmod +x ~/vox

Then add to ~/.claude.json:

{
  "mcpServers": {
    "vox": {
      "command": "/Users/$(whoami)/vox"
    }
  }
}

Restart Claude Code.

Option 3: Build from Source

git clone https://github.com/boska/vox
cd vox
swift build -c release --product vox

Add to ~/.claude.json:

{
  "mcpServers": {
    "vox": {
      "command": "/path/to/vox/.build/release/vox"
    }
  }
}

Restart Claude Code.

Voice

Default TTS voice is Hana via ElevenLabs (eleven_multilingual_v2). Supports Chinese, English, Czech, Vietnamese, and 20+ languages in the same voice. Without an ElevenLabs key, falls back to macOS system voice automatically.

How it works

Claude Code
    ↓ JSON-RPC 2.0 over stdio
  vox binary
    ↓                        ↑
AVAudioEngine        AVSpeechSynthesizer
SFSpeechRecognizer       ElevenLabs API
NLLanguageRecognizer

No ports. No sockets. No daemon. Just stdin/stdout.

Troubleshooting

"No speech detected" — speak within ~1s of calling listen, VAD cuts off after 0.8s silence.

No default input device — Mac mini has no built-in mic. Connect USB/Bluetooth mic or iPhone via Continuity Camera, set default in System Settings → Sound → Input.

ElevenLabs silent — add ELEVENLABS_API_KEY=sk-... to ~/.claude/.env, or leave it out to use system voice.