Image: KanaWorks AI
Why it mattersAn MCP server that runs an audio model on the developer's own machine turns asking an agent for a 10-second sound into a local file, not a bill.
Asking a coding agent for background music or a sound effect usually ends with a sign-up link or a plan the agent will not run. Audio Studio MCP, a new open-source server from KanaWorks AI, takes that step away. It runs Stable Audio 3 on the developer's own Mac or Windows PC, with no cloud service, no account and no API key, and Claude Code, Claude Desktop, Cursor or any other MCP client can call it in plain language.
The project is MIT-licensed and reached 183 stars on GitHub in its first day. It was published on 18 October 2026 and last pushed on 19 October.
What the server exposes
Four MCP tools make up the loop. generate_audio turns a prompt into 1 to 4 candidate clips of music, an instrument loop or a sound effect, each with its own seed. edit_audio has three modes: variation re-styles a whole clip, inpaint regenerates a time range, and continue extends a track. process_audio is a non-destructive chain that trims silence, resamples, loops seamlessly, applies fades, normalizes loudness or peak, and writes a 44.1 kHz 24-bit WAV, OGG or FLAC. unload_models frees the memory.
Every file that comes out has a <name>.json sidecar next to it, holding the prompt, the seed and every setting that produced it. On a Mac, the project's own README states that results for the same seed are bit-for-bit identical to the official sa3 CLI; on Windows it calls the official Stable Audio 3 PyTorch API.
The numbers KanaWorks published
The README gives timings on two machines. On an M1 Mac with 16 GB, the medium model generates 20 seconds of music in about 10 seconds once warm, two 3-second sound effects from a cold start in about 13 seconds, and continues a track to 35 seconds in about 27 seconds including the encoder load. On a Windows PC with an RTX 3060 12 GB, 20 seconds of warm music takes about 1.3 seconds and the maximum 380 seconds of music comes out in about 15 seconds. The small sm-music and sm-sfx models cover shorter clips (up to 120 seconds) at a fraction of the time.
Memory use is also stated. Loaded models hold 5 to 7 GB on a Mac. On the RTX 3060 the medium model peaked at about 5.3 GB of VRAM. Models unload after the idle timeout, which defaults to 10 minutes and is configurable with AUDIO_STUDIO_IDLE_UNLOAD_MIN.
What it is useful for
The README's own examples name a game and a scene for every call. Asking for a 60-second lo-fi loop at 85 BPM for a cafe scene with 3 candidates. A heavy wooden door slamming in a stone hallway, exported as OGG for Godot. Extending a track to 90 seconds and normalizing it to minus 16 LUFS, which the README says is a target for game background music.
The server also teaches the client the Stable Audio 3 prompt format, so a request like "make it sound like a creaky door" becomes the structured prompt the model was trained on (source, action, material, space). Without that prefix, sound effects often come out as music.
Audio is one of the places where asking an agent for help still breaks on billing and keys. A server that generates reproducible 44.1 kHz WAVs on the same machine the code sits on keeps the loop short and the data local.
Source
The code is at kanaworksai-hue/audio-studio-mcp on GitHub, under an MIT license.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
