macOS Apple Silicon (M1-M4) Required

Give your AI a Voice.
Keep the Control.

Bidirectional, interruptible voice loops for Claude, Cursor, and Gemini. Powered entirely by local ML on Apple Silicon. Zero API calls. Absolute privacy.

Press Shift to Talk
SHIFT
$brew install portaudio espeak-ng
$npm install -g voice-mcp-server

The Shift Advantage

We turned what others consider a limitation—Push-to-Talk—into a premium feature. Absolute control over your AI's context.

Flawless Focus

Immune to background noise. It only listens when you explicitly tell it to. No more false positives from keyboard clacking or barking dogs.

Hardware-Level Privacy

Your microphone isn't hot on your desk all day. The physical requirement to press Shift gives you total peace of mind.

A Self-Aware Audio Stack

The AI has meta-awareness. Ask it to 'switch to a more realistic voice' and it will autonomously hot-swap its own Python/Hydra backend on the fly.

Indefinite Standby Mode

Need time to think? Just tell the AI "give me a minute." It gracefully suspends itself, turns off the microphone lock, and waits silently until you press Shift again.

Voice-Native Pair Programming

It doesn't just speak; it collaborates. The AI intelligently manages the conversational floor, giving brief "muttering" updates while it works so you never feel abandoned.

Air-Gapped & Free

Your voice never leaves your machine. Powered locally by Apple MLX Whisper and Kokoro TTS. Code on an airplane, avoid per-minute API costs, and keep your microphone strictly private.

Surgical Interruption

True Barge-In

Interrupt mid-sentence with pinpoint precision. The moment you press Shift, the AI instantly stops speaking, listens to your new instruction, and logs the interruption.

It naturally understands it was cut off and pivots the conversation smoothly. No awkward overlapping audio.

AI:Here is the Python script you requested. First, we import the necessary libraries...
SYS:
{ "was_interrupted": true }
USR:Wait, actually write it in TypeScript instead.
AI:Oh, got it! Switching to TypeScript...

Firewall-Friendly Architecture

We explicitly engineered around the macOS firewall using local sockets. No annoying network popups. Deeply polished for local development.

Cost: $0.00Privacy: Absolute
MCP Client
(Claude/Cursor)
Node.js Bridge
(NPM)
SECURE LOCAL ENVIRONMENT
Python ML
Daemon
Whisper STT
Local MLX
Kokoro TTS
Local MLX
Optional Cloud API
ElevenLabs TTS
Remote Call

Setup in Seconds

Connect to your favorite MCP-compatible client effortlessly.

Input Monitoring Required

Because this uses a global Push-to-Talk (Shift) hotkey, you must grant Input Monitoring permissions to your terminal or client so it can detect the key globally.

System Settings > Privacy & Security > Input Monitoring > Toggle ON

The First Run

The first time you invoke the voice tool, it will silently download ~4GB of optimized local ML models (Whisper Large v3 & Kokoro).

Grab a coffee—your hardware-level privacy is worth the wait.
json
// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "voice-mcp-server": {
      "command": "voice-mcp-server",
      "args": []
    }
  }
}