Meet scorebench
scorebench is an agent-native desktop workbench for ScoreKit. You describe the music in natural language, the Agent turns that intent into a ScoreKit scene YAML file, and ScoreKit compiles and renders the YAML into playable, shippable audio.
Your brief
│
▼
scorebench Agent ──writes──► scene.yaml
│
▼
ScoreKit CLI
│
MIDI + renderer + FFmpeg
│
▼
OGG/WAV + meta.json + stems
│
▼
scorebench playback and spectrum
Three distinct roles
| Component | What it does | What it does not do |
|---|---|---|
| scorebench | Hosts the conversation, project files, Agent tool calls, read-only observation, playback, and spectrum views | It does not synthesize or mix audio, and it is not a piano-roll editor |
| Agent | Interprets the brief, arranges the music, writes scene YAML, and iterates on validation feedback | It does not bypass ScoreKit to generate audio directly |
| ScoreKit | Validates scenes, compiles deterministic MIDI, invokes renderers, and exports audio and metadata | It does not interpret natural language or decide what the piece should express |
The short version is: the Agent composes, ScoreKit compiles, and scorebench hosts the workflow.
Good fits
- Seamless game-music loops and related exploration, combat, or victory cues.
- Instrumental sketches for narrative or film-style scoring.
- Layered scores whose stems can be added or removed dynamically in a game engine.
- Projects that keep musical descriptions, scene files, and outputs under Git.
- Natural-language iteration with reproducible, inspectable build results.
Boundaries to understand first
- scorebench is not a DAW. It has no piano roll, timeline, or plugin chain.
- Scene YAML is an executable musical description, not an arbitrary prose container. Every field must match the live ScoreKit schema.
- The same scene produces stable MIDI. Final timbre still depends on the renderer, sound files, and external tool versions.
- ScoreKit and its sound sources are external dependencies. scorebench does not bundle ScoreKit, SoundFonts, or SFZ sample libraries into a project.
- The Review panel evaluates scene data, schema, validation results, semantic diffs, and render metadata. It has not listened to the audio; a human must still judge the sound.
If this is your first session, continue with Quick Start. If you can already render but need help translating musical intent into scene fields, go to From Protocol to Arrangement.