A sound studio your agent can actually call

What it does It’s an MCP server that turns plain-language requests into generated audio: BGM, instrument loops, and sound effects via Stable Audio 3, running locally on Apple Silicon (MLX) or an NVIDIA GPU (PyTorch). Beyond generation, it handles the boring-but-crucial finishing work — trimming, loudness normalization to LUFS, seamless looping, fades, and export to WAV/OGG/FLAC at 48 kHz. The demo shows Claude picking a loop length of exactly 11 bars so the beat stays in time on repeat, which tells you the agent is doing real musical reasoning, not just calling a function.

The interesting bit The server teaches the client the official Stable Audio 3 prompt format, so “make it sound like a creaky door” becomes a structured prompt the model actually understands. And on Mac, results for the same seed are bit-for-bit identical to the official sa3 CLI — reproducibility you rarely see in AI audio tooling.

Key highlights

• Three core tools: generate_audio (1–4 candidates per call), edit_audio (variation, inpaint, extend), and process_audio (a full finishing chain that never modifies the input)
• Every file gets a .json sidecar with prompt, seed, and settings
• unload_models frees GPU memory on demand
• Optional small, fast sm-music and sm-sfx models alongside the 1.4B medium model
• A sample page of unedited outputs, with prompts and seeds published

Caveats

• Not lightweight: ~6 GB of disk on Mac, ~14 GB on Windows, and 8 GB+ VRAM on the NVIDIA side
• Windows users need a Hugging Face account and license acceptance for the gated PyTorch weights — the installer handles this gracefully but it’s a real hurdle
• Linux isn’t mentioned at all; this is Mac-or-Windows only

Verdict Game developers and video creators who already live in Claude Code or Cursor get a genuinely useful local audio pipeline. If you’re on Linux or lack the hardware, move along.