Case Study
Giving AI assistants a voice, locally
The Cheema Text-to-Voice MCP Server lets any MCP-capable assistant speak: five built-in voices, four languages, and voice cloning, all running on your own machine. No API keys, no cloud, no per-character billing.
Why build it
Agents increasingly need audio output: reading results aloud, voice notifications, accessibility. The existing options meant cloud TTS with API keys, cost per character, and shipping your text to a third party.
The Model Context Protocol gives assistants a standard way to call tools, so I built the missing tool: a TTS server any MCP client can use, with inference running entirely locally.
How it works
The server is written in Python and exposes speech tools over three transports, so it plugs into whatever the client supports: stdio for desktop clients, SSE and HTTP for networked ones. Speech synthesis runs on NeuTTS with espeak-ng, including cloning a voice from a short reference sample.
- Five built-in voices across four languages
- Voice cloning from a reference recording
- stdio, SSE and HTTP transports
- Works with Claude Desktop, Claude Code, n8n, and any MCP client
- MIT licensed, listed on mcpservers.org
Where it lives
The server is published and listed on mcpservers.org, with source on GitHub under the MIT licence. Anyone can install it and give their assistant a voice in minutes.