The server is a bridge, not the synthesiser — AgentVRM has to be running before anything happens, and the README says so plainly with the endpoint it expects. What you get for that setup is more than text-to-speech: the avatar carries expressions and animation, and every generated line is written to disk as a .wav, so it can be reused in a video or a demo without another synthesis pass.
An MCP server that bridges a client to an AgentVRM engine. Text sent through it is synthesised into speech and performed by a VRM avatar with expressions and animation, rather than just being read aloud.
- Speech from text, performed by the VRM avatar, with a required `text` argument and optional `speaker_id`, `speed_scale` and `auto_play` — `speak_text`
- Automatic playback on generation, handled per platform: the default player on Windows, `afplay` on macOS, and `aplay` on Linux with a fallback to `xdg-open`
- The generated audio kept as a .wav file in the assets folder, so a line can be reused rather than regenerated
A running AgentVRM engine — it must be started manually, and the server connects to http://localhost:3001/api/speak_text by default, with `--api-url` to point elsewhere. Python 3.10 or higher. No install step with `uv`: run it straight through `uvx` as `agent-vrm-mcp`. The package is `agent-vrm-mcp` (0.1.0 in pyproject).
One command — uvx agent-vrm-mcp