SkillPartly tested (needs an account)
Compose original music with your agent
music-skill
Your assistant writes original instrumental music note by note, then renders, mixes and masters it to MP3 with MIDI and stems, with no AI audio model.
- Works with
- Claude Code, Codex CLI
- Good for
- Graphic designers
What it does
music-skill lets your assistant write original instrumental music the way a composer would. The assistant writes every note itself. The skill then checks the score for common mistakes, plays each part on recorded real instruments or built-in synthesizers, mixes the parts and masters the result.
No AI audio generator is used. Besides an MP3 and a WAV file you get the MIDI file (the notes themselves), one audio file per instrument (stems) and a project file. You can ask for a change to one bar, and only that part is played again.
It suits game music, short video soundtracks, loops for visual novels and similar work. Setup takes some time because the instrument sounds are large downloads.
Example prompts
- "Write a 90-second instrumental for the title screen of my puzzle game: warm, a little melancholic, piano and strings."
- "Write a looping cue for a sad scene in my visual novel: soft piano and cello, quiet enough to read over."
- "The chorus melody in bar 33 is too busy. Simplify it and render again."
- "What changed from the last version? Let me hear the difference."
- "Send me the MIDI and the stems too."
How to install
/plugin marketplace add owoevan-dev/music-skill/plugin install music-skill@music-skillnpx skills add owoevan-dev/music-skillHow to use it
First ask your assistant to set up the renderer and the sound banks (the README's Install commands). Then describe the music in plain words, for example a 90-second title theme with piano and strings.
Check it works
After setup, ask your assistant to run the self-check. SKILL.md gives the command: python <skill>/scripts/selfcheck.py --work <new empty folder> --config <config.json>. With a configuration it runs 136 checks, real renders included, in a few minutes.
Common problems
- macOS cannot write a game loop's Ogg copy (loop.ogg) — Homebrew's ffmpeg cannot do it; run brew install ffmpeg-full and pass its path with --ffmpeg to the bootstrap — https://github.com/owoevan-dev/music-skill
- Debian or Ubuntu cannot create the virtual environment — run sudo apt install python3-venv first — https://github.com/owoevan-dev/music-skill
- A project stops because a sound bank is missing — it prints the command that adds that bank — https://github.com/owoevan-dev/music-skill
Other assistants
- Codex and other agents: npx skills add owoevan-dev/music-skill, or copy skills/music-production into the agent's skills folder. The repo also ships a .codex-plugin manifest. The author reports no Codex test — https://github.com/owoevan-dev/music-skill
What to watch out for
- No paid API, no API key and no account. It uses only your assistant's normal model time
- Big downloads: all thirty sound banks are about 3.5 GB to download and 8 GB installed (keep about 12 GB free); Surge XT and its patches add about 270 MB. You can install only some banks with --only
- Render folders grow: the author measured 27 GB for one piece's folder. studio.py clean lists old renders and deletes them only with --remove
- Runs Python scripts, FluidSynth, FFmpeg and optionally Surge XT on your computer; setup uses pip and, on Linux, sudo apt install. A long piece can need up to about 2 GB of memory per synthesizer part
- Setup downloads tools and banks from GitHub (and the University of Iowa site when banks are built from source) and Python packages from PyPI, checked against pinned SHA-256 hashes. PRIVACY.md says it sends nothing: no telemetry, no account, no uploads
- Loads by itself whenever you ask for music, even just 'a song' (no disable-model-invocation)
- Code is MIT and surge_host.py is GPL-3.0-or-later. The project's banks are CC0 or free for any use; MS Basic, the fallback bank, is MIT and the README suggests you credit it. A bank you add keeps its own licence
- Young project: created 2026-11-02, version 0.4.0 released on 2026-11-08
Licence and author
- Licence
- MIT
- Author
- owoevan-dev
- Source code
- github.com/owoevan-dev/music-skill
How we tested it
We installed it on an empty test computer on 9 Nov 2026. Partly tested: installed in a clean test system, but its main features need an account or key we did not use.
The renderer, the sound banks and a render were not run. The README's /plugin lines are typed inside a Claude Code session. The author reports CI on Ubuntu and macOS and examples made with Claude; no Codex run is reported.
Technical details of the test
- Date
- 9 Nov 2026
- Method
- Reviewed by our team at commit c94b143 (v0.4.0): README, SKILL.md, PRIVACY.md, CHANGELOG, plugin manifests and the download code. In a clean test machine (Claude Code 2.1.296) the shell forms claude plugin marketplace add owoevan-dev/music-skill and claude plugin install music-skill@music-skill installed version 0.4.0.
Checked again on 9 Nov 2026
Alternatives
- gemini-api-cli — can make background music with Google's AI audio model and your Gemini API key; music-skill writes editable notes instead and needs no key
FAQ
- Q: Do I need a GPU or an account? A: No. The README says no GPU, no account and no audio model.
- Q: Can I use the music in my video? A: The README says nothing in the toolchain stands in the way: the project's banks need no credit, but credit MS Basic if a part fell back to it. A bank you add keeps its own licence. The example pieces are CC0.
- Q: Can the assistant hear the music? A: No. The skill's checks find problems, but a person has to listen to the result.
- Q: Does it work on Windows? A: Yes. On Windows x64 the bootstrap downloads pinned portable builds of FluidSynth and FFmpeg.
Sources
- github.com/owoevan-dev/music-skill
- github.com/owoevan-dev/music-skill/blob/c94b143/skills/music-production/SKILL.md
- github.com/owoevan-dev/music-skill/releases/tag/v0.4.0
- github.com/owoevan-dev/music-skill/blob/c94b143/CHANGELOG.md
- github.com/owoevan-dev/music-skill/blob/c94b143/PRIVACY.md