Free & Open Source

Give your agents a voice.

Text-to-speech plugin for Claude Code. 30 languages, voice design, voice cloning. Runs 100% locally on your Mac. Zero API calls, zero cost.

Get it on GitHub

Every voice in this demo was generated by Local TTS.

# add the marketplace

> /plugin marketplace add vdk888/local-tts

> /plugin install local-tts@local-tts-marketplace

# generate speech

> /local-tts:local-tts Say hello in French

✓ 30 languages · voice cloning · 100% offline

Three modes

Text-to-speech, voice design, and voice cloning

Most powerful

Voice Cloning

Clone any voice from a 3-10 second reference clip. Captures timbre, accent, and emotional tone. Layer style guidance on top for fine control. Ultimate mode with transcript for maximum fidelity.

INPUT reference audio (wav/mp3, 3-10s)

MODE reference cloning or ultimate (+ transcript)

STYLE combinable with voice design prefix

OUTPUT 48 kHz wav, cloned voice

Voice Design

Describe a voice in natural language — age, gender, accent, emotion, pace — and generate it. No reference audio needed. Works in any of the 30 supported languages.

Text-to-Speech

Simple, high-quality speech generation. Just pass your text and get a 48 kHz wav file. No configuration, no API keys. Auto-detects language from the text.

30 Languages

Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, and 15 more. No language tag needed.

100% Local

No cloud, no API calls, no cost. The model runs on your machine via PyTorch and Apple Metal (MPS). Your voice data never leaves your computer.

Installation

Up and running in two minutes

Add the marketplace, install the plugin, run setup. That's it.

Add marketplace

In Claude Code, run:
/plugin marketplace add vdk888/local-tts

Install plugin

Then install:
/plugin install local-tts@local-tts-marketplace

Auto-setup

First use creates a Python venv and downloads the model (~10 GB one-time). Takes ~5 minutes.

Start talking

Ask Claude to say something, narrate a text, or clone a voice. It just works.

Requirements

This plugin requires Claude Code and a Mac with Apple Silicon.

✓ macOS with Apple Silicon (M1/M2/M3/M4) ✓ Python 3.10+ ✓ ~10 GB free disk space ✓ Claude Code with plugin support

Linux with CUDA is also supported. Setup is automatic.

FAQ

Questions

Yes. The plugin is Apache-2.0 licensed. The underlying VoxCPM2 model is also Apache-2.0. No API keys, no subscriptions, no usage limits. It runs on your hardware, so the only cost is your electricity.

VoxCPM2 is a 2-billion parameter model. Quality is good for voiceovers, narration, and prototyping. It's not ElevenLabs-level, but it's free, offline, and runs instantly. Voice cloning with a good reference clip produces surprisingly natural results.

On Apple Silicon, expect ~2.3x real-time factor (10 seconds of audio takes ~23 seconds to generate). Model loading takes ~25 seconds per invocation. Great for batch work, not for live streaming.

Yes, if you have a CUDA-capable GPU. The setup script auto-detects your environment. On Linux with CUDA, generation is significantly faster than on Apple Silicon (torch.compile is available).

No. Everything runs locally. The model weights are downloaded once from Hugging Face, then everything runs offline. Your reference audio, generated speech, and text never leave your machine.

Bubble Invest. We run 20+ Claude Code agents in production for investment analysis, content creation, and infrastructure management. We built this plugin because we needed voice capabilities for our agents. The underlying model is VoxCPM2 by OpenBMB.