v0.2.1 — Duration parameter fix

A small bug-fix release. No license or breaking changes.

🎧 Listen to samples · English · 简体中文 · 日本語

Fixes

• generate_audio and edit_audio now accept seconds as well as duration_seconds for the output length. Before, a client that sent seconds had it silently ignored and got the default length instead (for example 60 s of music instead of 30 s). The tool schema still names the parameter duration_seconds, so existing calls work as before. Fixes #7, thanks to @peercaster for the report.

Update

From an existing clone:

git pull

Then restart your MCP client (or the server) so it loads the new version. Nothing needs to be reinstalled, and the model weights stay as they are.

New installs follow the README as before, using --branch v0.2.1.

Full changes: v0.2.0...v0.2.1

v0.2.0 — Windows (NVIDIA GPU) support

Audio Studio MCP now runs on Windows PCs with an NVIDIA GPU, as well as on Apple Silicon Macs. It is still all local: no cloud service, no API key.

🎧 Listen to samples · English · 简体中文 · 日本語

License

• The copyright holder in LICENSE and the authors field in pyproject.toml are now 東京AI映像合同会社 (Tokyo AI Visuals LLC). The license is still MIT.
• The KanaWorks_AI name, logo and the Kana character are now credited to 東京AI映像合同会社. As before, they are not covered by the MIT license (see docs/brand).

New: Windows (NVIDIA GPU)

• New sa3-torch backend built on the official Stable Audio 3 PyTorch API. It is used automatically on Windows and Linux; macOS keeps the MLX backend.
• New install.ps1 installer. It installs PyTorch with CUDA 12.6 (new cuda extra), clones the pinned Stable Audio 3 runtime and downloads the medium model (~10.4 GB). Use -Model all to also get the small sm-music / sm-sfx models.
• Full-length generations up to 380 s. Tested on an RTX 3060 12 GB, where the medium model peaked at about 5.3 GB of VRAM.
• The Windows weights are gated on Hugging Face. The installer asks for a read token when needed. When it runs without a console (for example, started by an AI agent), it does not prompt. It prints the steps to accept the model licenses and log in, then stops, and you re-run it afterwards. The READMEs include notes for AI agents that install it for you.
• New AUDIO_STUDIO_DEVICE setting picks the PyTorch device (cuda, cuda:1, cpu).
• .json sidecars and other text files are now written as UTF-8, so non-ASCII prompts work with any Windows locale.
• Linux with an NVIDIA GPU should work with the same backend, but it is untested and has no installer yet (see the FAQ).

Fixes (macOS)

• On Apple Silicon, install.sh could pick an x86_64 Python running under Rosetta (for example an old Intel Homebrew). That silently skipped MLX and later failed with "Intel Macs are not supported". The installer now asks uv for an arm64 Python, rebuilds a non-arm64 .venv, and checks after install that .venv is arm64 and that MLX imports. It also explains a failed uv sync when UV_PYTHON is set. The server now tells you when it is running under Rosetta. Fixes #3, thanks to @HirokiSatoyoshi for the report and diagnosis.

Docs

• The READMEs (English, 简体中文, 日本語) now cover Windows: requirements, install, configuration, performance and FAQ.
• The banners and the social preview now show Mac and Windows.
• The commercial-use FAQ now says "less than US$1M in total annual revenue", matching the Stability AI Community License.
• Added a star request to the READMEs.

Install

Let your agent do it. Paste into Claude Code:

Install https://github.com/kanaworksai-hue/audio-studio-mcp by following its README, then register it as an MCP server.

Or by hand. macOS (Apple Silicon):

git clone --branch v0.2.0 https://github.com/kanaworksai-hue/audio-studio-mcp.git
cd audio-studio-mcp
./install.sh

Windows (NVIDIA GPU):

git clone --branch v0.2.0 https://github.com/kanaworksai-hue/audio-studio-mcp.git
cd audio-studio-mcp
powershell -ExecutionPolicy Bypass -File install.ps1

No release assets are needed. The installers fetch the Stable Audio 3 runtime (pinned at 77e0398) and the model weights from Hugging Face.

Requirements

• macOS: Apple Silicon Mac (M1 or later), 16 GB RAM recommended, about 6 GB of disk
• Windows: 64-bit Windows 10 or 11, an NVIDIA GPU with 8 GB of VRAM or more and a current driver, about 14 GB of disk, and a Hugging Face account that has accepted the Stable Audio 3 medium license
• Python 3.10–3.13, git, ffmpeg

Full changes: v0.1.0...v0.2.0

v0.1.0 — First public release

First public release of Audio Studio MCP: local Stable Audio 3 music and sound effects for Claude Code, Claude Desktop, Cursor and any other MCP client. Everything runs on your own Mac. No cloud service, no API key.

🎧 Listen to samples · English · 简体中文 · 日本語

What's in it

• generate_audio: text to BGM, instrument loops/stems or sound effects, 1–4 candidates per call
• edit_audio: variation, inpaint (regenerate a time range) and continue (extend a track)
• process_audio: trim, seamless loops, fades, loudness (LUFS), 48 kHz WAV / OGG / FLAC
• unload_models: free model memory right away
• A .json sidecar with prompt, seed and settings next to every file
• Built-in Stable Audio 3 prompt guidance for the client
• Bit-for-bit identical output to the official sa3 CLI for the same seed

Install

Let your agent do it. Paste into Claude Code:

Install https://github.com/kanaworksai-hue/audio-studio-mcp by following its README, then register it as an MCP server.

Or by hand:

git clone --branch v0.1.0 https://github.com/kanaworksai-hue/audio-studio-mcp.git
cd audio-studio-mcp
./install.sh

The installer fetches the Stable Audio 3 runtime (pinned at 77e0398) and the medium model (~5.9 GB) from Hugging Face. No release assets are needed.

Requirements

• Apple Silicon Mac (M1 or later), 16 GB RAM recommended
• Python 3.10–3.13, git, ffmpeg

Windows support is on the roadmap.