bkmeneguello/glados-voice-mcp ? reverse-engineered prompt

Build me a small Python project that lets AI agents talk through a shared GLaDOS voice on my machine, exposed as an MCP tool.

I want it to be easy to install and run with a simple command, then register in an MCP client. It should start a local background voice service the first time someone speaks, load the models once, and let multiple agents share the same audio device without talking over each other. If something is already speaking, return a busy message and ask the caller to try again later. Also make it so sending empty text stops the current speech for everyone.

Please include the voice models download flow, a basic command to test speaking, and a Python API I can use directly too. It should support a CPU setup by default, with an option for GPU if available. Keep it local only, with no browser access, and make sure the prompt for the MCP tool makes it clear the text has to stay in character.