mcp-soundfx

Local text-to-SFX generation for AI agents. An MCP server that runs Stability AI's Stable Audio Open 1.0 on your own NVIDIA GPU, so any MCP-capable agent (Claude Code, Claude Desktop, etc.) can generate broadcast-quality 44.1 kHz stereo WAV sound effects from text prompts — fully offline after the initial model download, no API costs, no usage limits.

Built and used in production at Core Epoch to generate shipping game audio: UI clicks, inhale/exhale foley, ambient loops, and achievement stingers, all authored by an agent iterating on prompts and auditioning seeds.

What it does

Exposes one tool, generate_sound:

The tool's docstring embeds a full prompt-engineering and negative-prompt guide, so the agent reads how to write good audio prompts without any human in the loop. The model excels at SFX, foley, ambience, and musical phrases; it does not generate speech or vocals.

VRAM-friendly by design: the pipeline loads in float16 (~2.5 GB VRAM), generates, and is torn down after every call — so it coexists with games, renderers, or anything else that owns the GPU between generations.

Related MCP server: Freesound MCP Server

Setup

Requirements: Windows/Linux, Python 3.10+, NVIDIA GPU with ~3 GB free VRAM, CUDA-capable torch.

1. Accept the gated model license (one time): log into Hugging Face and accept the terms on the stable-audio-open-1.0 page, then create a Read token at Token Settings.

2. Install:

   cd mcp-soundfx
   python -m venv .venv
   .venv\Scripts\activate          # Windows (source .venv/bin/activate on Linux)
   pip install torch --index-url https://download.pytorch.org/whl/cu121
   pip install -r requirements.txt

3. Register the server with your MCP client. For Claude Code / Claude Desktop:

   {
     "mcpServers": {
       "mcp-soundfx": {
         "command": "powershell.exe",
         "args": ["-ExecutionPolicy", "Bypass", "-File", "C:\\path\\to\\mcp-soundfx\\run.ps1"],
         "env": { "HF_TOKEN": "hf_your_read_token" }
       }
     }
   }

   The launch scripts use .venv next to the script by default; set MCP_SOUNDFX_VENV to point elsewhere.

4. Smoke test (optional, ~30 s on first run after model download):

   python test_generation.py

The first generation downloads the model weights (~2.5 GB) to your Hugging Face cache; everything after that is offline.

Why an MCP server instead of a script?

Because the interesting workflow is agentic: "give the pause menu a soft airy whoosh, try three variations, keep the least clicky one" is a conversation, not a command line. Putting the generator behind MCP means the agent owns the loop — prompt, generate, listen (via whatever playback tool it has), re-prompt — and sound design becomes something you delegate rather than operate.

License

The server code is MIT © 2026 Core Epoch LLC.

The model (Stable Audio Open 1.0) is gated and separately licensed by Stability AI under the Stability AI Community License — free for research, non-commercial, and commercial use by entities under $1M annual revenue; you accept those terms directly with Stability AI when you unlock the model, and generated outputs are yours. This project is not affiliated with or endorsed by Stability AI.