Cadence Code

Natural voice conversations with Claude Code, Codex, and Antigravity, fully local on Apple Silicon.

Talk through a bug, redirect a task, or hear a quick update when your hands are busy or your eyes need a break. Your coding agent still does the thinking; Cadence Code simply handles speech.

• Local speech: TTS and STT run on your Mac with MLX.
• Natural turns: speak, listen, interrupt, and continue without reloading models.
• Explicit control: Cadence Code listens only after you start a conversation.
• No extra reasoning layer: no daemon, cloud speech service, or second language model.

Quick start

Requires an Apple Silicon Mac running macOS 14+, Python 3.11-3.14, a microphone, and an output device. Internet access is needed for installation and the first model downloads.

Claude Code

/plugin marketplace add michael-L-i/cadence-code
/plugin install cadence-code@cadence-code-marketplace

Restart Claude Code, then run:

/cadence-code:start-talking

Codex

codex plugin marketplace add michael-L-i/cadence-code
codex plugin add cadence-code@cadence-code-marketplace

Start a new Codex session, then run $start-talking or choose Start Talking from /skills.

Codex plugins work in the CLI and desktop app. The IDE extension does not yet support them.

Antigravity

git clone https://github.com/michael-L-i/cadence-code
agy plugin install ./cadence-code

Restart Antigravity, then run /start-talking.

On first run, Cadence Code selects recommended defaults, asks for microphone access, and downloads the speech models automatically. If Claude Code is still finishing the one-time setup, run /reload-plugins when prompted and start again; other hosts just need a restart.

Talking with Cadence Code

Action	Claude Code	Codex	Antigravity
Start a conversation	/cadence-code:start-talking	$start-talking	/start-talking
Interrupt and redirect	/cadence-code:jump-in	$jump-in	/jump-in
Change speech models	/cadence-code:voice-settings	$voice-settings	/voice-settings
End and release models	/cadence-code:wrap-up	$wrap-up	/wrap-up

You can also say "stop" or "goodbye" to end a conversation.

How it works

Cadence Code runs as a lightweight stdio MCP server beside your coding agent. It loads local speech models only when a conversation starts, keeps them warm between turns, and releases them when the conversation ends.

Raw audio and speech inference stay on your Mac. Your transcript is returned to your coding agent, which decides what to do and composes the exact response spoken back to you. See Cadence Code privacy for the complete boundary.

Only one Cadence Code conversation can use the microphone and model memory at a time.

Voice models and download sizes

Text to speech

Model	Tier	Language	Download
Pocket TTS 100M	Lightweight	English	236 MB
Kokoro 82M	Lightweight	English	389 MB
Chatterbox Turbo 350M	Balanced	English	417 MB
Qwen 0.6B	Highest quality	English	1.97 GB

Speech to text

Model	Tier	Language	Download
Moonshine Base 61M	Lightweight	English	248 MB
Parakeet 110M	Balanced	English	459 MB
Parakeet 0.6B v3	Highest accuracy	25 languages	2.51 GB

Pocket TTS and Parakeet 110M are the defaults. TTS is currently English-only; choose Parakeet 0.6B v3 for multilingual transcription. Follow each linked model card for its license terms.

Configuration

Use the voice settings command to choose models. Advanced model, voice, speed, endpointing, and audio-device settings live in config.toml:

• Codex, Antigravity, and direct development: ~/.cadence-code/config.toml
• Claude Code: the plugin's data directory

Model weights use the shared Hugging Face cache at ~/.cache/huggingface/hub. Depending on your choices, downloads require about 500 MB to 4.5 GB.

Troubleshooting

• No microphone or output: allow mic access for your coding agent in System Settings > Privacy & Security > Microphone, check the configured audio devices, and restart the host.
• Setup or download fails: check your Python version, internet connection, and free disk space, then restart the host to retry.
• Session already in use: stop Cadence Code in every Codex, Claude Code, Antigravity, and development session.
• An update still shows the old version: fully exit every host process that loaded Cadence Code, then start a new one.

For checkout-based diagnostics:

uv run --locked cadence-code doctor
uv run --locked cadence-code listen-test

Updating and uninstalling

To update, rerun the marketplace and plugin update commands for your host, then fully restart it. A running MCP process is not replaced in place.

To uninstall:

# Codex
codex plugin remove cadence-code@cadence-code-marketplace
codex plugin marketplace remove cadence-code-marketplace

# Claude Code
claude plugin uninstall cadence-code@cadence-code-marketplace
claude plugin marketplace remove cadence-code-marketplace

# Antigravity
agy plugin uninstall cadence-code

Antigravity updates the same way: pull or re-clone the repository and rerun agy plugin install.

Configuration, private environments, and shared model weights may remain after uninstalling. Codex and Antigravity data lives in ~/.cadence-code; model weights live in the Hugging Face cache.

Development

uv.lock is the source of truth for local development and CI.

uv sync --locked --python 3.13
uv run --locked cadence-code doctor
./dev check    # tests + plugin validation
./dev claude   # local branch test in Claude Code
./dev codex    # local branch test in Codex
./dev agy      # local branch test in Antigravity

See AGENTS.md for the project map, CONTRIBUTING.md before changing dependencies, and RELEASING.md for release details.

Community

Questions and ideas are welcome in GitHub Discussions. Use the issue forms for bugs and feature proposals.

Cadence Code is released under the MIT License. Please follow the Code of Conduct, and report security issues privately as described in SECURITY.md.