audio-studio-mcp is an open source MCP server that runs Stable Audio 3 on your own machine, on Apple Silicon (MLX) or Windows NVIDIA GPUs (PyTorch). It makes background music and sound effects for your videos with no cloud and no API key. It works in Claude Code and any MCP client, so your agent can produce audio locally.
Claude Code (or any MCP client)
|
| generate_music / generate_sfx
v
audio-studio-mcp (local MCP server)
|
+-- Apple Silicon --> MLX
+-- Windows NVIDIA --> PyTorch
|
v
music.wav / sfx.wav on your disk
|
v
editor + Creator OS (post to 11 platforms)What is audio-studio-mcp and how does local Stable Audio work?
audio-studio-mcp turns a text prompt into an audio file on your own hardware, using Stable Audio 3 as the local model.
The idea is simple. Most music and sound effect tools are a request to someone else’s server. audio-studio-mcp runs the model next to your agent. When Claude Code asks for a track, the server loads Stable Audio 3, generates the audio, and writes a file. Nothing leaves the machine. There is no account to create and no key to paste. The project is on GitHub at kanaworksai-hue/audio-studio-mcp, which is also where install steps and flags live.
Because it speaks MCP, the audio generator behaves like any other tool in your agent’s toolbox. The same session that schedules posts can also produce a 20 second bed of music for a Reel.
How to install audio-studio-mcp in Claude Code (Apple Silicon and NVIDIA)
Add the server to Claude Code’s MCP config, then restart Claude Code so it picks up the new tools.
The exact install command and repo URL are in the project’s README. The shape is the same as any MCP server entry: a name, a command, and arguments. Two hardware paths matter:
| Machine | Runtime | What to expect |
|---|---|---|
| Apple Silicon Mac (M series) | MLX | Runs on the GPU cores through MLX, no CUDA, no separate driver install |
| Windows PC with NVIDIA GPU | PyTorch | Runs on the CUDA GPU, needs a working NVIDIA driver and PyTorch build |
After the restart, ask Claude Code what audio tools it has. Once the server is listed, you can prompt for a track in plain language. Keep the model files and generated audio on a disk with room to spare.
How to generate background music and SFX for short-form video with one prompt
One prompt is enough to get a usable music bed, and a second short prompt gets you a sound effect to accent the cut.
Here is a realistic session in Claude Code:
Generate a 25 second background music track for a vertical
short. Warm lo-fi piano, soft vinyl crackle, no vocals,
steady tempo, ends cleanly so it can loop under a voiceover.
Save it as bgm_lo_fi.wav.
Then generate three sound effects: a soft whoosh for a
transition, a short pop for a text reveal, and a subtle
riser for the final call to action.Ask for length, mood, instruments, and whether vocals are allowed. For short-form, tell the model the track sits under a voice, so it should not fight the mid range. For SFX, describe the action, not the file: “a whoosh that sounds like a page turning in a hurry” gets a better result than “whoosh.wav”.
Keep a small library: one bed per tone (calm, energetic, tutorial), and a handful of transitions. Reuse them across videos so your channel sounds consistent. Watch the walkthroughs on the KevBuildsApps YouTube channel for short-form editing patterns that pair well with generated audio.
Can local AI audio replace licensed music and stock SFX?
For background beds and everyday transitions, generated audio covers most short-form needs. For a signature track or a famous sound, it does not.
Generated music is best at the layer nobody listens to on purpose: the bed under a talking head, the loop under a product montage, the riser before a reveal. It is not a replacement for a composer’s theme or a licensed song people already recognize. If your video depends on a specific track, you still need a license.
- Use generated audio for: background beds, transitions, stingers, UI pops, ambience, placeholder tracks during editing.
- Keep licensing for: recognizable songs, artist collaborations, brand anthems, anything a viewer would Shazam.
One practical benefit: you can regenerate. If the bed is too busy, ask for a sparser version. Stock libraries give you one file.
audio-studio-mcp vs cloud music APIs: cost, speed, and copyright risk
The trade is upfront hardware and setup against per generation cost and a network round trip, with a shorter path to knowing what you own.
| audio-studio-mcp (local) | Cloud music API (general) | |
|---|---|---|
| Cost model | Your hardware and electricity, no per track fee | Typically metered per generation or subscription |
| Speed | Depends on your GPU; no upload or queue wait | Network round trip plus provider queue |
| Privacy | Prompts and audio stay on your machine | Prompts and outputs pass through a provider |
| Licensing | Check the model and project license before publishing | Check the provider’s terms per plan |
| Offline | Works without an internet connection | Requires a connection |
On copyright: local generation removes the “did I license this track” question at the source, but it does not remove the need to read the model’s license and your platform’s rules. If you publish to YouTube or TikTok, treat generated audio like any other asset: keep the prompt and the file, and know what the license allows. Read the project README for the model terms before you publish.
How to plug generated audio into your Creator OS video pipeline
Generate the audio locally, mux it into the video in your editor, then let Creator OS publish the finished file across platforms.
Creator OS is one place to publish, schedule, reply and read analytics on 11 platforms plus Skool, and it ships a hosted MCP server and a CLI so an agent can run the accounts. That means the same Claude Code session can generate a track with audio-studio-mcp and then schedule the finished video through Creator OS. The video editor skills in OPEN VIDEO EDIT handle the edit side: captions, b-roll cards, and cuts, rendered locally with Python and ffmpeg. Add audio after the picture is locked, so the bed matches the final length.
A workable loop:
- Edit the clip and lock the length. See the one line launch video skill.
- Generate the music bed and two or three SFX with audio-studio-mcp.
- Mux audio in your editor, then export the final file.
- Publish or schedule through Creator OS from the CLI or the MCP server, for example:
creatoros media:upload final_short.mp4
creatoros posts:create --text "New short is live" \
--platforms instagram,tiktok,youtube \
--media med_abc123 \
--scheduledAt 2026-03-02T17:00:00Z --timezone America/TorontoFor the API side and per platform video rules, start at the Creator OS docs. The content repurposing agent walkthrough shows how a longer video becomes several shorts in the same session. If you want an agent that runs the whole social loop on its own, Social Agents is the open source harness, and the launch video covers how it fits together.
FAQ
Does audio-studio-mcp need an API key or internet connection?
No. The project description says no cloud and no API key, because Stable Audio 3 runs on your own machine. Once the model files are downloaded, generation does not depend on a network connection. Check the README for the first run, which may need to fetch model weights.
What hardware do I need to run Stable Audio 3 locally?
Apple Silicon for the MLX path, or a Windows machine with an NVIDIA GPU for the PyTorch path. Those are the two backends the project names. Expect model files and generated audio to use disk space, and expect generation time to depend on the GPU.
Is AI-generated music from a local model safe to use on YouTube and TikTok?
Local generation removes the license question at the source, but it does not replace reading the model license and the platform rules. Keep the prompt and the output file with your project. For anything where a specific song matters to the video, use a license instead.
How long does it take to generate a 30-second music track locally?
It depends on your GPU and the model settings, and the project does not publish a fixed benchmark, so time it once on your own machine. Run a single 20 to 30 second test before you plan a batch. After that, generation runs without a queue or network wait.
Can audio-studio-mcp run inside any MCP client, not just Claude Code?
Yes. The project describes itself as local AI background music and sound effects for Claude Code and any MCP client. If your client can add a local MCP server, it can call the same tools. See the Creator OS MCP docs for how a hosted MCP server connects in different clients.
Get started
Generate the audio on your machine, then publish the finished video without juggling five dashboards. Creator OS plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included. Create an account at https://www.creatoros.ca/sign-up, and watch the walkthroughs on the KevBuildsApps YouTube channel.
Get started for $0
Free trial for new subscribers. Cancel anytime.