Agent Audio is an MIT-licensed Python project that gives local coding agents audio generation through a Model Context Protocol (MCP) server and an Agent Skill. It installs its own runtime and uses Stable Audio 3 Medium to generate audio on the user's computer, without ComfyUI or a paid audio-generation API. The idea is that an agent working on a game or video can handle useful sound effects as part of the task, and the developer reviews the finished result instead of managing a separate audio workflow. The README labels the project an early preview (v0.1): short audio generation has been demonstrated in the listed environments, while automatic use during game and video work and listening quality still need end-to-end validation. Setup is meant to be done by pasting a prompt into a coding agent, which checks Git, Python 3.11+ and uv, installs the runtime and models, and registers the MCP server and Skill for Codex, Claude Code and Cursor; manual setup uses uv and install/bootstrap.py. The MCP server exposes two tools, audio_status and generate_audio, which writes a WAV file and returns its path. Duration must be greater than zero and at most 380 seconds, subject to a 540-second inference timeout. On Windows the default backend is TFLite/LiteRT on CPU, with model downloads of about 7.36 GB and 25 GB or more of free disk recommended; on Apple Silicon an MLX backend is used, with 6.88 GB of model files and a short generation reported on a 24 GiB Mac. The interface is text-to-WAV only, not audio-to-audio editing, inpainting or continuation, and nonempty negative prompts are rejected. The source code is MIT licensed, but model weights are distributed separately under Stability AI's Community License, with T5Gemma components subject to Gemma terms, so users must review and accept those terms themselves. The README describes it as an integration tool rather than a newly trained model.