Rheo MCP Server

The Name — Rheo (ῥέω)

"Everything flows, nothing stands still." — Heraclitus

Rheo is derived from the ancient Greek verb ῥέω (rhéō), meaning "to flow", "to stream", or "current" — the root of rheology and immortalized in the philosophy of panta rhei (πάντα ῥεῖ, "everything flows").

In speech synthesis and vocal interaction, voice is never static; it is a continuous, living acoustic stream. The project was christened Rheo to capture this fluid essence: the frictionless flow of human thought into synthetic voice, and spoken word into zero-latency dictation. Rheo embodies an uninterrupted, private stream of vocal intelligence running entirely under your own sovereignty.

Related MCP server: leanvox-mcp

Overview

Rheo is an open-source, local-first voice computing environment engineered for maximum fidelity, complete privacy, and zero data leakage. By executing state-of-the-art neural speech models locally on modern GPUs and Apple Silicon, Rheo delivers studio-grade voice cloning and responsive speech-to-text without recurring API subscriptions or cloud latency.

Whether you need to generate expressive audiobooks, clone your vocal identity for content creation, dictate continuously into native desktop applications, or empower AI coding agents with speech via the Model Context Protocol (MCP), Rheo provides a unified and extensible platform.

Key Capabilities

• Zero-Shot Voice Cloning: Clone reference voices using short audio samples (3–10 seconds) with exceptional timbre and prosody retention.

• Multi-Engine Speech Synthesis: Seamlessly swap between leading open-weights models tailored for latency, multilingual coverage, or ultra-expressive acting.

• Global System Dictation: Hold a global keyboard chord to capture system microphone audio and transcribe it directly into whichever application is currently focused.

• Model Context Protocol (MCP) Integration: Expose your voice catalog directly to Claude, Cursor, and custom agent workflows with native tool calls.

• Local Timeline & Multi-Track Studio: Compose, arrange, splice, and export speech segments with customizable paralinguistic tags, pitch control, and audio effect chains.

• 100% Offline by Design: Weights are downloaded once from HuggingFace and execute entirely in memory on your workstation. No analytics, tracking, or telemetry.

Supported Neural Engines

Quickstart

Prerequisites

• Node/Bun: Bun >=1.0.0 or Node.js >=20.0

• Python: Python 3.11, 3.12, or 3.13

• Rust: Rust toolchain 1.75+ (for building the Tauri desktop bundle)

• GPU Tooling:

  • NVIDIA: CUDA 12.1+ / cuDNN

  • AMD: ROCm 6.0+

  • Apple: macOS 12.0+ (Metal Performance Shaders / MLX)

Installation & Development

# Clone repository
git clone https://github.com/MasterZ1311/Rheo.git
cd Rheo

# Install workspace frontend dependencies
bun install

# Configure Python virtual environment & backend packages
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\Activate.ps1
pip install -r backend/requirements.txt

# Run development server with Tauri desktop shell
bun run dev

Running Backend Independently

uvicorn backend.main:app --host 127.0.0.1 --port 17493 --reload

Hardware Support

Rheo automatically inspects your compute topology on initialization and dynamically assigns optimal tensor backends:

• NVIDIA GPUs: Native PyTorch with Tensor Cores, flash-attention, and CUDA execution.

• AMD Radeon GPUs: Native ROCm acceleration on Linux and Windows (RDNA 2, RDNA 3, and RDNA 4).

• Apple Silicon: Metal Acceleration via mlx and mlx-audio unified memory architecture.

• CPU Fallback: Optimized AVX-512 / AVX2 execution paths using lightweight models like Kokoro and LuxTTS.

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    Rheo Desktop Shell                       │
│              (Tauri v2 + Rust Core Engine)                  │
│                                                             │
│   ┌─────────────────────────────────────────────────────┐   │
│   │               React Workspace UI                    │   │
│   │      (Vite + Tailwind CSS + TanStack Query)         │   │
│   └──────────────────────────┬──────────────────────────┘   │
└──────────────────────────────┼──────────────────────────────┘
                               │ HTTP / WebSocket (Port 17493)
┌──────────────────────────────▼──────────────────────────────┐
│                    Rheo Neural Backend                      │
│                  (FastAPI + Async Python)                   │
│                                                             │
│   ┌───────────────────────┐     ┌───────────────────────┐   │
│   │     TTS Backends      │     │      STT Backend      │   │
│   │ (Qwen3, Chatterbox,   │     │   (Whisper PyTorch /  │   │
│   │  TADA, Kokoro, Lux)   │     │       MLX Engine)     │   │
│   └───────────────────────┘     └───────────────────────┘   │
│   ┌───────────────────────┐     ┌───────────────────────┐   │
│   │      SQLite DB        │     │      MCP Server       │   │
│   │  (Profiles & Studio)  │     │  (AI Agent Protocols) │   │
│   └───────────────────────┘     └───────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Project Watermark & Ownership

• Project: Rheo

• Repository: MasterZ1311/Rheo

• Maintainer: MasterZ1311 (thenappanmasterz1311@gmail.com)

• License: GNU Affero General Public License v3.0 (AGPL-3.0-or-later)