name: music-agent-workflows
description: "Configure and operate Muzic MusicAgent safely: install
prerequisites, manage model and cache layout, validate config, launch CLI or
Gradio, and troubleshoot tool selection and secrets."
metadata:
disco-role: operating
disable-model-invocation: true
license: MIT
Music Agent Workflows
Use this sub-skill when the request is about MusicAgent itself: setup, configuration, model and download layout, startup, tool enablement, or runtime troubleshooting.
Start here
What it covers
- Docker vs Conda/pip installation trade-offs.
- Linux system packages needed for audio, notation, and model download support.
config.yaml fields, local_fold, src_fold, log_file, and secret handling.
- CLI vs Gradio launch paths and their credential differences.
- Task planning, tool selection, and model/tool enablement boundaries.
- Safe config validation without importing the source repo.
What it does not cover
- ROC, DiffSinger, DDSP, and other tool internals.
- Generic provider or gateway policies outside MusicAgent.
- Training or algorithm-level guidance for the wrapped models.
Operating notes
- Keep API keys, tokens, and
.env contents out of version control.
- Run MusicAgent from the project directory that contains
config.yaml so relative paths resolve as expected.
- Not every planner task has a loaded pipe. Some task names are aspirational and will fall through to an "unloaded models" error until the plugin layer is extended.
- If a task fails, check
disabled_tools, then the local_fold cache, then the required system package or credential.
- Use the bundled validator before trying to launch the agent.