Scoreproof

scoreproof

Convert any score. Prove nothing got lost.

Scoreproof converts music files between MusicXML, MIDI, ABC, MEI and Humdrum, makes sheet music you can print and play (PDF, PNG), repairs the quirks of real notation-app exports (Sibelius, Dorico, Finale, MuseScore, Guitar Pro and others), and checks every conversion note by note, reporting exactly what was kept and what was lost.

MIDI you played into a DAW or from a keyboard comes out as readable notation: Scoreproof finds the beat (even without a click), snaps rhythms to 16ths or triplets, tidies note lengths, splits piano into two hands and spells notes in the right key.

Use it from any AI tool that speaks MCP (Claude, Cursor, VS Code, Gemini CLI, Codex and others), from the command line, or from Python.

Use it in claude.ai, the Claude apps or ChatGPT

No install. Add a custom connector with this address:

https://scoreproof-fkkuctjokq-uc.a.run.app/mcp

• Claude (web, desktop, mobile): Settings → Connectors → Add custom connector.
• ChatGPT: Settings → Connectors → Create (developer mode).

Then ask "Write a short waltz and give me the sheet music". To work on your own file, drop it on the Scoreproof page and paste the link it gives you into the chat; the same page converts files with one click.

The hosted service is free, has per-address rate limits and a 10 MB file limit, and deletes every file after 24 hours.

Use it from your AI tool

Scoreproof runs locally and works on files on your computer. It needs uv (brew install uv or pip install uv).

Claude Code

/plugin install scoreproof --marketplace asampat3090/scoreproof

or add the server directly: claude mcp add scoreproof -- uvx scoreproof mcp

Claude Desktop, Cursor, Gemini CLI: add this to claude_desktop_config.json, ~/.cursor/mcp.json or ~/.gemini/settings.json:

{

"mcpServers": { "scoreproof": { "command": "uvx", "args": ["scoreproof", "mcp"] } } }

VS Code (.vscode/mcp.json):

{

"servers": { "scoreproof": { "type": "stdio", "command": "uvx", "args": ["scoreproof", "mcp"] } } }

Codex (~/.codex/config.toml):

[mcp_servers.scoreproof]

command = "uvx" args = ["scoreproof", "mcp"]

Then ask things like:

• "Turn ~/Music/take-3.mid into sheet music I can print."
• "Convert ~/Music/song.mid to MusicXML and tell me what got lost."
• "Transpose my lead sheet to Bb and show me the sheet music."
• "Split this score into parts for each player."
• "What changed between draft-1.musicxml and draft-2.musicxml?"
• "Check this Sibelius export before I send it to the copyist."

Tools

Results are written next to the input under a new name. Scoreproof never overwrites an existing file unless you give that exact output path.

Use it from the command line

uvx scoreproof convert song.musicxml song.mid

uvx scoreproof convert take-3.mid --to pdf # played MIDI to sheet music uvx scoreproof inspect song.musicxml --bars 1-8 uvx scoreproof transpose song.musicxml --key "Bb major" uvx scoreproof extract score.mxl --each-part uvx scoreproof render song.musicxml --to png --page-size a4 uvx scoreproof --help

Add --json to any command for machine-readable output.

Use it from Python

from scoreproof import convert

result = convert("song.musicxml", "song.mid") result.confidence # "exact", "high", "medium" or "low" result.expected_losses # what MIDI cannot hold: lyrics, chord symbols, ... result.fixes # exporter quirks repaired on the way in

How accurate is it?

On 430 real exports from 11 notation apps (benchmarks/score_corpus/):

MIDI played in a DAW (benchmarks/played_midi/, five pieces performed with human timing):

Every conversion re-reads its own output and compares it with the source. When a format cannot hold something (MIDI has no chord symbols or lyrics layout; ABC and MEI exports carry transposing instruments at concert pitch), the report says so instead of failing silently.

Not supported: native Sibelius, Dorico, Finale and MuseScore files. Export MusicXML from the app first.

Development

pip install -e ".[dev]"

pytest --cov=scoreproof # coverage is kept at 100% black src tests benchmarks flake8 src tests benchmarks

The plan and milestones are in docs/plan.md. MIT licensed.