Levity

Voice-driven AI development — talk to your coding assistant and hear it talk back.

Levity is a hands-free voice layer for AI-assisted development. It captures your voice, transcribes it locally with Whisper, and speaks responses aloud using a two-tier text-to-speech engine (a fast local system voice, with an optional cloud upgrade via Gemini 2.5 Flash TTS).

The project ships two independent components — use either or both:

Component	What it is	Use it when
`levity-voice-mcp`	An MCP server that gives Claude Desktop, Antigravity IDE, and standalone Google Antigravity.app (Agentic) a voice.	You want to talk to your AI assistants hands-free.
`antigravity-voice`	A VS Code / Antigravity extension with its own STT, TTS, wake-word, and settings UI.	You want voice control inside your editor.

Both share the same design: local-first, bring-your-own-key (BYOK), no telemetry.

Platform support

The levity-voice-mcp server is cross-platform. It auto-detects the OS and uses the native speech engine on each:

Platform	Status	Speech engine
macOS	✅ Supported	`say` + `afplay` (built in)
Windows 11	✅ Supported	PowerShell `System.Speech` (SAPI) + `System.Media.SoundPlayer` (built in)
Linux	✅ Supported	`espeak-ng`/`espeak` (`sudo apt install espeak-ng`) + `aplay`/`paplay`/`ffplay`

Speech-to-text for voice_confirm uses Whisper, which runs on all three. The older antigravity-voice VS Code extension is still macOS-focused; the cross-platform support above applies to the MCP server.

Prerequisites

Shared across both components:

Python 3.9+ on your PATH (python3 --version).
A working microphone, and permission for your terminal / editor / Claude Desktop to access it (macOS: System Settings → Privacy & Security → Microphone).
~150 MB–1.5 GB disk for the Whisper model (downloaded on first run; size depends on the model you choose — base is ~140 MB).
Xcode Command Line Tools on macOS (xcode-select --install) — needed to build some audio dependencies.

Optional, depending on features you enable:

Gemini API key (free tier available at aistudio.google.com) — enables the higher-quality cloud TTS voice and, in the extension, AI command processing. Without it, everything falls back to the local system voice.
Picovoice access key (free at console.picovoice.ai) — only the antigravity-voice extension's wake-word mode needs this. The MCP server has no wake-word; it's TTS plus on-demand voice_confirm.
A speech engine on Linux — sudo apt install espeak-ng. macOS and Windows use built-in engines.

1. `levity-voice-mcp` (Claude Desktop, Antigravity IDE, and Standalone Antigravity.app)

A lightweight MCP server exposing voice tools to Claude Desktop, the Antigravity IDE, and the standalone Google Antigravity.app (Agentic):

voice_speak — speak text aloud (native system voice, or Gemini 2.5 Flash TTS for longer replies). Returns immediately; audio plays in the background.
voice_confirm — quick spoken yes/no (≤5s, Whisper STT) → yes/no/unclear for hands-free approvals (e.g. "should I run this?"). Only proceed on yes.
voice_listen — full free-form listening (up to 30s, Whisper STT) → returns the raw transcript, so you can answer anything ("let's go with option A").
voice_toggle — start/stop the server, toggle voice responses, set listen mode, check status.

Response modes. A listen_mode preference (quick or full) — toggled from the menu bar's Response Mode item — tells the assistant which style you prefer: quick favors voice_confirm (yes/no), full favors voice_listen (free-form).

Setup and Deployment

Step 1: Install & Create Virtual Environment

One-Click Install (macOS)

In Finder, open the levity-voice-mcp/ folder and double-click install.command. This will automatically create your local virtual environment, install dependencies, copy the server scripts to a stable path at ~/.levity-voice/, and register the server inside your Claude Desktop configuration.

Manual Setup (macOS/Linux Scripted)

Or, perform a manual installation in your terminal:

cd levity-voice-mcp
./setup.sh          # creates ~/.levity-voice/venv and installs deps

Note: This script copies server.py and menubar.py into your home directory under ~/.levity-voice/ for stable referencing.

Step 2: Register the MCP Server

Depending on which environment(s) you use, copy the following configuration block into the corresponding config file:

A. Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "levity-voice": {
      "command": "/Users/YOUR_USERNAME/.levity-voice/venv/bin/python",
      "args": ["/Users/YOUR_USERNAME/.levity-voice/server.py"]
    }
  }
}

B. Antigravity IDE

Add to ~/.gemini/antigravity-ide/mcp_config.json:

{
  "mcpServers": {
    "levity-voice": {
      "command": "/Users/YOUR_USERNAME/.levity-voice/venv/bin/python",
      "args": ["/Users/YOUR_USERNAME/.levity-voice/server.py"]
    }
  }
}

C. Standalone Google Antigravity.app (Agentic)

Add to ~/.gemini/antigravity/mcp_config.json:

{
  "mcpServers": {
    "levity-voice": {
      "command": "/Users/YOUR_USERNAME/.levity-voice/venv/bin/python",
      "args": ["/Users/YOUR_USERNAME/.levity-voice/server.py"]
    }
  }
}

(Make sure to replace YOUR_USERNAME with your actual macOS/Linux account name).

Step 3: Run and Control the System

Once registered, restart your respective application (Claude Desktop, Antigravity IDE, or Antigravity.app).

Interactive macOS Menu-Bar App

The server launches the menu bar automatically (macOS) — whenever the voice server starts (host launch, or when you start the voice service) the 🎙 Levity icon appears. A single-instance guard prevents duplicate icons across Claude Desktop / Antigravity. Disable with "auto_menubar": false in ~/.levity-voice/config.json.

To start it by hand (e.g. before the server runs):

~/.levity-voice/venv/bin/python ~/.levity-voice/menubar.py

For a Login Item, click "Launch at Login" in the menu. (An optional double-click app can be built with build-app.command, but the automatic launch above is the recommended path.)

This shows the Levity status-bar icon at the top-right of your screen:

Response Mode — choose Quick (Yes/No) or Full (Free-form) input.
Server: ON/OFF — start or stop the voice server.
Voice Response: ON/OFF — mute/unmute spoken responses.
Restart Server — reload the server in place (picks up updated server.py).

Direct Agent Prompts

You can also direct your AI assistant to control the voice engine by typing or saying:

"Start the voice server."
"Mute voice responses."
"Ask me to confirm before running that." (uses voice_confirm)

Configuration

Settings live in ~/.levity-voice/config.json (created on first run). Notable keys:

Key	Default	Meaning
`whisper_model`	`base`	`tiny` / `base` / `small` / `medium` — accuracy vs speed (used by `voice_confirm`/`voice_listen`).
`listen_mode`	`quick`	`quick` (yes/no) or `full` (free-form). Set from the menu bar.
`local_voice`	platform default	System voice name. macOS: `say -v '?'`; Windows: SAPI voice name; Linux: espeak voice.
`gemini_api_key`	`""`	Set to enable cloud TTS (or use the `GEMINI_API_KEY` env var / `~/.levity-voice/.env`).
`gemini_voice`	`Kore`	Gemini TTS persona.

Config changes are hot-reloaded — the running server picks them up without a restart.

2. `antigravity-voice` (VS Code extension)

A full voice assistant inside the editor: Listen → Transcribe → Think → Speak. With a Gemini key it answers coding questions with awareness of your active file; without one it echoes back your transcription.

Install (development build)

cd antigravity-voice
npm install
npm run compile        # builds ./out
npm run setup-sidecar  # creates sidecar/.venv and installs Python deps

Then press F5 in VS Code to launch an Extension Development Host, or package a .vsix with vsce:

npx vsce package

Usage

Hotkey: Cmd+Alt+V (macOS) / Ctrl+Alt+V (Windows/Linux) to start/stop.
Command Palette: search "Antigravity" for all commands.
Settings: run "Antigravity: Open Voice Settings" for a visual config panel, or edit antigravity.* keys in VS Code settings.
API keys: "Antigravity: Set Gemini API Key" / "Set Picovoice Access Key" — stored securely in the OS keychain via VS Code SecretStorage, never on disk.

Trigger modes

Tap-to-talk — press the hotkey to start, press again (or pause) to stop.
Wake word — always-on detection; say the keyword to start (needs a Picovoice key).

Wake word

Wake-word activation lives only in the antigravity-voice extension, via Picovoice Porcupine (free access key, supports custom .ppn keywords). The levity-voice-mcp server is intentionally TTS plus on-demand voice_confirm — hands-free input there comes from your assistant's own voice input or voice_confirm, not an always-on wake word.

Voice input & dictation

Levity captures your spoken answers with its own Whisper engine (voice_confirm / voice_listen) — it does not depend on any third-party dictation app (e.g. Dictaflow) or on macOS Dictation. Nothing extra is required for Levity's voice features to work.

If you also like to dictate your typed chat messages, that's separate from Levity: use whatever you prefer. If you don't have a tool like Dictaflow, macOS Dictation is built in (System Settings → Keyboard → Dictation) and works out of the box — Levity neither requires nor manages it.

Guaranteed voice output

You can make Levity speak every completed response, not just when the model remembers to. How strong that guarantee is depends on the runtime, because a true guarantee has to run when a turn ends — outside the model's control.

Runtime	Mechanism	Guarantee
Claude Code / Cowork	`Stop` hook (`levity-voice-mcp/hooks/`) — cross-platform	✅ Every turn, unless muted or no TTS engine
Antigravity extension	in-code `TTSProvider.speak()` after each turn	✅ Every turn, unless muted
Claude Desktop	model calls `voice_speak`	⚠️ Best-effort — Desktop has no hook system

Claude Code / Cowork: install the Stop hook (see levity-voice-mcp/hooks/README.md). It reads the final assistant message and speaks it deterministically. A last_spoken.json marker prevents double-speak when the model also voiced the turn.
Claude Desktop: there is no post-turn hook, so spoken output relies on the model calling voice_speak (the server's tool instructions strongly enforce this). voice_speak now returns immediately and plays audio in the background, so long replies no longer time out and silently drop.
Mute anytime: say "voice off" (runs voice_toggle response_off) or set response_active: false in ~/.levity-voice/config.json. Every mechanism honors this switch.

Skill & Cowork plugin

Skill (levity-voice-mcp/skill/SKILL.md) — lets the assistant start and drive the voice loop on command ("start voice"). Installed for Antigravity under ~/.gemini/config/skills and global_skills.
Cowork plugin (cowork-plugin/levity-voice/) — packages that skill for Claude Desktop's plugin system; build a .plugin to install it there. It's skill-only (it uses your existing levity-voice MCP rather than adding a second server).

Levity currently runs as a single active voice server — one host at a time (a PID lock prevents multiple processes fighting over the mic). Running it across multiple hosts simultaneously is designed in docs/multi-host-voice-daemon.md.

Privacy & security

Speech-to-text runs locally (Whisper) — your audio is not sent anywhere unless you explicitly enable cloud TTS.
API keys are stored in the OS keychain (extension) or your local config file (MCP server) — never committed to the repo and never logged.
No analytics or telemetry.

Platform notes

The MCP server detects the OS at startup (platform.system()) and routes TTS accordingly — no configuration needed:

Windows: speech via PowerShell System.Speech; all subprocesses launch with CREATE_NO_WINDOW so no console flashes. Self-restart uses Popen + os._exit (since os.execv doesn't replace in place on Windows).
Linux: install a speech engine — sudo apt install espeak-ng — and an audio player (aplay/paplay/ffplay, usually already present).
macOS: works out of the box.

The older antigravity-voice VS Code extension has not yet been ported and remains macOS-focused.

Roadmap

Multi-host voice daemon — run Levity in Claude Desktop + Antigravity simultaneously via a shared daemon + per-host shims (design).
Port the antigravity-voice extension to the cross-platform engine.
Fully offline AI tier (local LLM via Ollama for command interpretation).
Pinned dependency versions for reproducible installs.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Levity

Platform support

Prerequisites

1. `levity-voice-mcp` (Claude Desktop, Antigravity IDE, and Standalone Antigravity.app)

Setup and Deployment

Step 1: Install & Create Virtual Environment

One-Click Install (macOS)

Manual Setup (macOS/Linux Scripted)

Step 2: Register the MCP Server

A. Claude Desktop

B. Antigravity IDE

C. Standalone Google Antigravity.app (Agentic)

Step 3: Run and Control the System

Interactive macOS Menu-Bar App

Direct Agent Prompts

Configuration

2. `antigravity-voice` (VS Code extension)

Install (development build)

Usage

Trigger modes

Wake word

Voice input & dictation

Guaranteed voice output

Skill & Cowork plugin

Privacy & security

Platform notes

Roadmap

License

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 8 Commits
antigravity-voice		antigravity-voice
cowork-plugin/levity-voice		cowork-plugin/levity-voice
docs		docs
levity-voice-mcp		levity-voice-mcp
.gitignore		.gitignore
LICENSE		LICENSE
Levity Blueprint.md		Levity Blueprint.md
README.md		README.md
build-app.command		build-app.command

Folders and files

Latest commit

History

Repository files navigation

Levity

Platform support

Prerequisites

1. levity-voice-mcp (Claude Desktop, Antigravity IDE, and Standalone Antigravity.app)

Setup and Deployment

Step 1: Install & Create Virtual Environment

One-Click Install (macOS)

Manual Setup (macOS/Linux Scripted)

Step 2: Register the MCP Server

A. Claude Desktop

B. Antigravity IDE

C. Standalone Google Antigravity.app (Agentic)

Step 3: Run and Control the System

Interactive macOS Menu-Bar App

Direct Agent Prompts

Configuration

2. antigravity-voice (VS Code extension)

Install (development build)

Usage

Trigger modes

Wake word

Voice input & dictation

Guaranteed voice output

Skill & Cowork plugin

Privacy & security

Platform notes

Roadmap

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

1. `levity-voice-mcp` (Claude Desktop, Antigravity IDE, and Standalone Antigravity.app)

2. `antigravity-voice` (VS Code extension)

Packages