# RVC Voice MCP

Turn text from an MCP-capable chat client into speech, convert it with a local RVC `.pth + .index` model, and render an audio player in the conversation.

> This project generates synthetic audio. Only use voices you own or have permission to use, and clearly disclose AI-generated output.

## How it works

1. ChatGPT calls `speak_rvc` with text.
2. The local server asks Applio to synthesize a base voice and run RVC conversion.
3. Depending on `PLAYBACK_MODE`, the local computer plays the WAV, the MCP Apps widget renders it, or both happen.

The server never accepts model paths or shell commands from chat. Paths are fixed in local environment configuration and excluded from Git.

## Requirements

- Windows and Node.js 20+
- A working Applio installation, either inside `local/Applio` or at a manually configured path
- An RVC inference `.pth` and matching `.index`
- ChatGPT Developer Mode and a public HTTPS tunnel for web integration

## Local setup

```powershell
Copy-Item .env.example .env
npm install
npm run check
npm start
```

Edit `.env` before starting and replace `HTTP_TOKEN` with a long random secret.

## Local automatic discovery

The repository reserves these ignored local directories:

```text
local/Applio/    Complete Applio installation
local/models/    RVC .pth and .index files
```

If the four path settings in `.env` are left empty, the server uses `local/Applio/env/python.exe` and searches `local/models` recursively. Automatic model selection is enabled only when exactly one `.pth` and exactly one `.index` file are present, preventing an accidental mismatch when several voices are installed.

Manual paths still take priority. You can set any or all of these values independently:

```env
APPLIO_DIR=D:\Apps\Applio
PYTHON_PATH=D:\Apps\Applio\env\python.exe
RVC_MODEL_PATH=D:\Voices\character.pth
RVC_INDEX_PATH=D:\Voices\character.index
```

Restart the service after adding or changing models because discovery happens at startup.

## Playback modes

Choose where generated speech plays:

```env
PLAYBACK_MODE=widget
```

- `widget` plays through the ChatGPT audio card.
- `local` starts playback through the computer running this server.
- `both` enables local playback and keeps the ChatGPT player available.

On Windows, local playback works without another dependency. To use ffplay instead, set its full path:

```env
LOCAL_PLAYER_PATH=C:\ffmpeg\bin\ffplay.exe
```

- Health: `http://127.0.0.1:8788/health`
- MCP: `http://127.0.0.1:8788/mcp?token=YOUR_TOKEN`

## One-click public endpoint

Run:

```powershell
npm run public
```

The first run downloads Cloudflare Tunnel into the ignored local `tools` folder. The script generates a fresh access token, starts the MCP server and tunnel in the background, then prints the HTTPS connector URL to add in ChatGPT Developer Mode. Keep that terminal open only while reading the URL; the two background processes continue running.

Stop both processes with:

```powershell
npm run stop
```

Quick tunnels are intended for development and their URL changes each time. For a stable public release, configure a named Cloudflare Tunnel and a domain you control.

## Tools

- `speak_rvc(text)` generates AI speech and renders the player.
- `get_status()` checks local configuration without exposing paths.

## Current scope

The initial adapter uses Applio's built-in TTS command and RVC inference. Fully offline TTS and a stable named tunnel are future work.

## License

MIT