What is Audio Studio MCP?
Audio Studio MCP lets Claude Code, Claude Desktop, Cursor or any MCP client ask for a piece of background music, an instrument loop or a sound effect in plain English and get finished files back. Generation runs locally on Stability AI's Stable Audio 3, through MLX on Apple Silicon or PyTorch on an NVIDIA GPU under Windows, so there is no API key and no per-clip charge.
It exposes four tools: generate_audio (one to four candidates per call), edit_audio (re-style a clip, regenerate a time range, or extend a track), process_audio (trim, resample, normalise loudness, make a gapless loop, fade and export to WAV, OGG or FLAC) and unload_models. A local web UI on the same port offers the same functions with a library of generated files, and each file gets a JSON record of its prompt and seed.
It is a one-person project from KanaWorks_AI under the MIT licence, created on 18 October 2026; v1.0.0 shipped on 24 October 2026. The model weights are under the Stability AI Community License, not MIT.
How to install Audio Studio MCP
git clone https://github.com/kanaworksai-hue/audio-studio-mcp && ./install.shIs Audio Studio MCP worth it?
Audio Studio MCP is more than a thin wrapper around Stable Audio 3 because of what happens after generation. process_audio does the dull work that usually sends you to an audio editor: gapless loops with a crossfade, loudness normalised to a LUFS target (the README suggests -16 for game music and -14 for streaming), resampling to 48 kHz, fades and export to OGG for a game engine. The demo shows Claude choosing a loop length of exactly 11 bars so the beat stays in time when it repeats, which is the kind of detail a prompt-only tool leaves to you.
The hardware bar is real. On a Mac it needs Apple Silicon, 16 GB of RAM is recommended, and the medium model is about 5.9 GB on disk and uses 5 to 7 GB of memory once loaded. On Windows it needs an NVIDIA GPU with 8 GB of VRAM or more and about 14 GB of disk, and the PyTorch weights are gated behind a Hugging Face login and licence click. Linux should work with the same backend but has no installer and is untested; Intel Macs and AMD GPUs are not supported.
Speed is reasonable on both. The README reports about 10 seconds for 20 seconds of music on an M1 with 16 GB, and about 1.3 seconds for the same on an RTX 3060 12 GB, where the maximum 380-second track takes about 15 seconds warm. The small sm-music model returns 15 seconds of music in about a second. Stable Audio 3 cannot sing, so there are no vocals or lyrics, and the README advises prompts of one or two English sentences.
The licence needs reading before commercial use. The code is MIT, but the weights are under the Stability AI Community License, which is free for commercial use only for organisations under US$1M in annual revenue, counting all revenue. The web UI binds to 127.0.0.1 with no login, and the README warns against exposing it, since anyone who reaches it can use your GPU and write files. With one contributor, six days of commits and none since 24 October, it is a promising v1.0 rather than a dependency to build a studio pipeline on.
Where it earns its place
A solo indie developer building a Godot game in Claude Code
Get a loopable lo-fi track and a set of UI sound effects without leaving the editor
Ask for 60 seconds of 85 BPM lo-fi as three candidates, pick one, and have process_audio cut a seamless OGG loop at -16 LUFS into the project folder. Costs nothing beyond the GPU time; the Community License covers a studio under US$1M revenue.
A YouTube editor at a three-person channel on an M-series Mac
Replace stock-library background music with tracks that fit each video's length
Generate a 30-second bed, extend it with edit_audio continue to the cut's length and normalise to -14 LUFS for streaming. About 10 seconds per 20 seconds of music on an M1.
A sound designer prototyping effects for a mobile game
Try many variations of an impact or ambience quickly
Fill source, action, material and space fields in generate_audio, ask for four variations per call and keep the seed in the JSON sidecar to reproduce the one that works.
Before you commit
- Setup
- Clone the repository and run ./install.sh on macOS or install.ps1 on Windows; the script installs uv, the pinned Stable Audio 3 runtime and the medium weights (about 5.9 GB on macOS, 10.4 GB on Windows), then prints the claude mcp add command. Windows needs a Hugging Face login with the licence accepted; ffmpeg is required on both.
- Integrations
- Four MCP tools over stdio, or over HTTP at /mcp when started with --http or --web; documented for Claude Code, Claude Desktop and Cursor. A local web UI in English, Japanese and Chinese shares the same model and queue.
- Data and privacy
- Prompts and audio stay on your machine; nothing is uploaded. Files go to outputs/ or AUDIO_STUDIO_OUTPUT_DIR with a JSON record. The web UI has no login and listens on 127.0.0.1 only. Code is MIT; weights follow the Stability AI Community License.
- Cost at scale
- No usage fees. The real cost is hardware: an Apple Silicon Mac with 16 GB recommended, or an NVIDIA card with at least 8 GB VRAM (the medium model peaked at about 5.3 GB on an RTX 3060). Companies over US$1M annual revenue need a paid Stability AI Enterprise licence, price not published.
- Limits
- Medium model up to 380 seconds, small models up to 120 seconds; instrumental only; English prompts advised; no LoRA support yet; Linux untested, Intel Macs and AMD GPUs unsupported.
What’s good
- Local generation with no API key or per-clip cost
- Seamless loops, LUFS loudness and OGG export in one tool
- Re-style, inpaint and extend existing clips
- Web UI and MCP share one process and one loaded model
What’s not
- Apple Silicon or an 8 GB NVIDIA GPU required; Linux untested
- No vocals; Stable Audio 3 is instrumental only
- Weights need an Enterprise licence above US$1M revenue
- One maintainer; created 18 October 2026
Best for
Indie game developers and video makers with an Apple Silicon Mac or NVIDIA PC who want an agent to produce loopable music and effects without a subscription.
Audio Studio MCP alternatives
Audio Studio MCP: common questions
- Is Audio Studio MCP worth it?
- A well-finished local sound studio for game and video makers on a recent Mac or NVIDIA card, more useful for its looping and loudness tools than for the generation alone. We rate it 3.7 out of 5.
- How much does Audio Studio MCP cost?
- Free under the MIT licence; Stable Audio 3 weights free for commercial use under US$1M annual revenue, Enterprise licence from Stability AI above that.
- What are the downsides of Audio Studio MCP?
- Apple Silicon or an 8 GB NVIDIA GPU required; Linux untested; no vocals; Stable Audio 3 is instrumental only; weights need an Enterprise licence above US$1M revenue; one maintainer; created 18 October 2026.
- Who is Audio Studio MCP best for?
- Indie game developers and video makers with an Apple Silicon Mac or NVIDIA PC who want an agent to produce loopable music and effects without a subscription.
- How much does Audio Studio MCP cost for a team or in production?
- No usage fees. The real cost is hardware: an Apple Silicon Mac with 16 GB recommended, or an NVIDIA card with at least 8 GB VRAM (the medium model peaked at about 5.3 GB on an RTX 3060). Companies over US$1M annual revenue need a paid Stability AI Enterprise licence, price not published.
- Is Audio Studio MCP safe for company data?
- Prompts and audio stay on your machine; nothing is uploaded. Files go to outputs/ or AUDIO_STUDIO_OUTPUT_DIR with a JSON record. The web UI has no login and listens on 127.0.0.1 only. Code is MIT; weights follow the Stability AI Community License.
- What are the best Audio Studio MCP alternatives?
- The closest alternatives we have reviewed are ElevenLabs MCP.
Answers checked Nov 5, 2026.