Skip to content
music.build

music.build is an MCP server that lets AI agents compose music as typed, immutable data.

Introduction

music.build is an MCP server that lets AI agents compose music by building typed, immutable structures — notes, voices, harmony, and formal sections — and exporting them to MIDI and LilyPond notation.

There is no piano roll, no GUI, and no drag-and-drop. Compositions emerge entirely from tool calls.

Why typed music data?

Music is hierarchical: a note belongs to a phrase, a phrase to a section, a section to a piece. When those relationships are explicit and typed, an AI agent can reason about them — transposing a voice, checking voice leading, assembling an AABA form — without ever losing precision.

In music.build, every value is exact:

  • Durations are rational fractions (3/8, not 0.375) — no floating-point drift
  • Pitches carry their spelling (F#4 is distinct from Gb4) — enharmonics are preserved
  • Transforms return new values; originals are untouched — no side effects

This makes compositions replayable, queryable, and structurally editable at any point in the session.

How it works

The server exposes 47 MCP tools across seven areas:

AreaWhat it does
VoicesBuild and edit named note sequences
HarmonySet keys, chord progressions, generate walking bass and comp voices
TransformsTranspose, invert, retrograde, augment — all produce new voices
FormDefine sections, repeats, and volta endings; assemble into a final score
DrumsLoad named drum patterns onto MIDI channel 9
RulesCheck voice leading, meter, and instrument ranges
ExportWrite MIDI, LilyPond PDF, and a session log to disk

An AI agent connects to the server, builds up a composition voice by voice, and calls export.all when done. The output lands in a numbered folder under generated_tracks/.

The data model

The central unit is a Voice — an ordered sequence of note events. Events are typed: Note, Rest, or Chord. Voices accumulate in a CompositionContext (one per server process), along with instrument assignments, dynamics, and form state.

When the agent calls export, the context assembles a Score — the complete composition — and hands it to the MIDI and LilyPond renderers.

css
Note / Rest / Chord   ← note events
       ↓
     Voice            ← named sequence of events
       ↓
  CompositionContext  ← session state (voices, parts, form)
       ↓
     Score            ← assembled composition
       ↓
  MIDI + LilyPond     ← output files

What agents can compose

With music.build connected to Claude or another MCP-capable agent, the agent can:

  • Build note sequences with full pitch spelling, rhythm, and articulation
  • Layer voices into a score and assign GM instruments
  • Generate walking bass lines, chord voicings, and diatonic harmonizations
  • Assemble multi-section AABA or verse-chorus forms with volta endings
  • Check voice leading and instrument range rules before export
  • Export to MIDI and, if LilyPond is installed, engraved PDF sheet music

Next steps

  • Quickstart — compose a piece in five minutes
  • Installation — Java 25 setup and Claude Desktop config
  • Tools — full reference for all 47 MCP tools
Edit this page →