earakely-scale/MuseScoreFromRequestComputerEnv-REA

MuseScore 3: a computer-use environment built from a one-line request

The real MuseScore 3.2 music notation editor running on an Ubuntu 24.04 desktop, as an agentenv computer environment. An agent completes the task only through screenshots, mouse and keyboard, and a verifier grades the saved files with no model.

Claude Code built this environment from a one-line request in about 32 minutes, with only REA's build skill and MCP tools; How it was made has the request and the steps. MuseScore opens a four-bar practice piece, Documents/MuseScore3/Scores/Morning_Song.mscz, at launch, behind its first-run Startup Wizard and Tour. The agent has to change the title shown on the score from Morning Song to Evening Lullaby and the composer from Theo Ashby to Clara Hollis, and save. It sees only the screen and acts only with the mouse and keyboard.

Watch

• Agent runs (Claude Code with Claude Sonnet 5.5): run 1, run 2, run 3
• Witness, which proves the task can be done through the computer: witness.mp4

Each video is the computer's own screen recording at 10 frames per second.

The task

In MuseScore, change the title of the score in Documents/MuseScore3/Scores/Morning_Song.mscz to Evening Lullaby and its composer to Clara Hollis, then save it.

bundle/artifacts/retitle-score-check/verify.py reads the app's files through the computer's data/get view and passes only if every check holds:

Check	Kind	File
title	text_contains	Documents/MuseScore3/Scores/Morning_Song.mscz
composer	text_contains	Documents/MuseScore3/Scores/Morning_Song.mscz
title-frame	text_equals	Documents/MuseScore3/Scores/Morning_Song.mscz

An .mscz file is a zip holding the score as .mscx XML, and REA reads the entry inside the zip. The title and composer are Text elements in the score's title frame (VBox), which have no attributes to select by, so the checks read the frame's whole text: it must contain "Evening Lullaby" and "Clara Hollis", and equal "10TitleEvening LullabySubtitlePractice piece in C majorComposerClara Hollis" exactly, the frame's height and each line's style and text run together. The exact match also proves the subtitle is untouched and nothing was added. Changing only the title in File > Score Properties leaves the title frame alone and fails.

Results

3 of 3 runs passed, about $0.86 per run.

Run	Outcome	Turns	Cost	Video
1	passed	17	$1.12	video
2	passed	18	$0.87	video
3	passed	18	$0.58	video

Run it

You need Docker. agentenv-framework builds the computer: Ubuntu 24.04 with the app installed from Ubuntu's archive (musescore3 3.2.3).

pip install agentenv-framework==0.9.1431
hf download earakely-scale/MuseScoreFromRequestComputerEnv-REA --repo-type dataset --local-dir musescore-request
agent-env run musescore-request/bundle --task retitle-score --agent claude-code --model <model>

The agent needs a model provider configured for agentenv. --task smoke only boots the computer, with no agent. bundle/README.md describes the bundle in detail.

How it was made

Claude Code built this environment from a one-line request, given only REA's build skill and REA's MCP tools:

Build me a MuseScore environment where the agent changes the title and composer of a short score and saves it.

It asked only the skill's upfront questions, wrote no app-specific scripts, and finished in about 32 minutes.

It took a second capture. On the first, REA did not count the score as MuseScore's document: MuseScore's window title shows the score's title rather than its file name, so the export refused a check on the score. Claude Code started the second capture with the score's path as MuseScore's launch argument, saved the seed at that path, and the export accepted it. The steps were REA's usual ones:

1. Capture. REA launched MuseScore 3 on a Linux desktop and recorded its package, windows and opened files while the edit was made. It also recorded the app's state folders before and after.
2. Export. REA turned the capture into an agentenv computer_app. It installs the app, seeds the starting files, starts the app, and waits for its main window. It also serves typed views of the app's files for verifiers.
3. Task. The task's checks were validated against the capture: they fail on the starting state and pass on the captured end state.
4. Witness. In a fresh computer, the edit was made through computer actions alone, after dismissing the Startup Wizard and Tour, and the checks were read at four stages. They held on the first try:

Stage	Checks
seed	fail, as expected
after_edit	pass
near_miss	fail, as expected: the composer was retyped as "Clara Hollins"
after_reset	fail, as expected: data/reset restored the seed

The agent runs came after the build: Claude Code with Claude Sonnet 5.5 (--model anthropic/claude-sonnet-5-5 --no-emulate, on an amd64 host) solved the task 3 times, each run recorded.

What's inside

bundle/      the agentenv bundle: envs/ (app module, seed files), tasks/, artifacts/ (verifier), evals/
media/       witness and agent-run recordings (MP4) with short GIF previews
results.json the runs above, with each check's result

Limits

• MuseScore's settings are not part of the seed, so every boot shows its five-page Startup Wizard and then a Tour dialog, which the agent has to dismiss. After data/reset, MuseScore also asks whether to restore the previous session.
• The exact title-frame check fails if the agent adds or restyles a line in the title frame, even when the title and composer read correctly.
• The environment is named mscore3, after the program the musescore3 package installs, so its files are under bundle/envs/mscore3/.
• The computer installs whatever musescore3 Ubuntu 24.04's archive holds at build time (3.2.3 for these runs); it is not pinned.
• Three runs is a small sample: the solve rate shows the task is solvable, not how hard it is.

Licences

The bundle code, the task and the synthetic seed files are Apache-2.0. MuseScore 3 itself is not included: the computer installs it from Ubuntu's archive (musescore3 3.2.3) at build time, under its own licence (GPL-2.0 with MuseScore's font exception).