Why We Built Open Notation Tools for AI Agents

A month ago I argued in this Journal that AI cannot replace the composer — that a generated score is not composition any more than a calculator is mathematics. I still believe that. Nothing about the position has softened.

And yet — every week, more AI agents in the world are being asked to write music for somebody. A teacher asks for a four-bar example to illustrate a half cadence. A novelist asks for the song their character is humming. A parent asks the agent to render the lullaby they made up for their child. The agents oblige, because that is what they are built to do.

What they produce is the part of the work that matters least — and they often produce it badly. MIDI you cannot read. Lead sheets that drift across bar lines. Whole notes labeled as quarters because no one taught the agent the difference. Engraving that looks like it came from a fax machine. Theory explanations that confidently invent a chord progression Bach would never have written.

> The short answer: AI agents are going to render music whether or not anyone helps them. We chose to help them — by handing over the engraving press and the music-theory reference, not the composer's voice. We trust the craft to remain with the human composer. We do not trust the rendering pipeline to remain accurate without intervention.

So we built something. A free, open music notation API and a small piece of software that any AI agent can install in one line. Send a score in JSON, get back inline SVG with proper engraving (the same software IMSLP and the Music Encoding Initiative use), real MusicXML that opens cleanly in Sibelius and Dorico, and a MIDI file that plays the right notes. Pre-flight validation that tells the agent exactly how its input is malformed. A small theory-knowledge endpoint that returns curated chunks from our curriculum and our chorale analyses — so when an agent looks up deceptive cadence before it generates a progression, it reads what Boulanger taught, not what a training corpus interpolated.

This is not a product launch. It is a position.

The principle

We are not giving AI the composer's voice. We are giving AI the engraving press.

The composer's voice is the accumulated residue of a specific human life — every piece they have written, every score they have studied, every performance they have heard. It cannot be extracted from a model and it cannot be transferred to one. The piece that no training corpus contains, because it does not yet exist, is still going to come from a human composer.

The engraving press is something else. It is the typesetter who turns the manuscript into a clean, readable score. It is the music-theory reference book on the desk. It is the kind reviewer who tells you that your Mm. 7 is in the wrong meter. None of those things are composition. They are infrastructure. They have always been provided by skilled craftsmen who were not the composer.

When the engraving is wrong, students learn the wrong thing. When the theory is hallucinated, the answer is plausible and false. When a generated chorale has parallel fifths the model never noticed, somebody copies them into their homework and gets confused when the teacher marks it. The damage is not large — but it is consistent, and it accumulates. Better engraving in the world is a public good. Better theory references in the world are a public good.

The conservatory tradition has always provided these things. We are simply extending the practice to include the agents.

What this is not

This tool does not generate music. It renders the notation an agent has already decided on. The composition — the actual selection of notes, the shaping of the line, the harmonic argument — happens upstream, in whatever model the agent is running. We do not bless that composition. We do not validate that it is good music. We render it accurately, surface concrete errors when they exist, and refer the agent to authoritative theory sources when it asks. The judgment of whether the music is any good still requires a human ear.

This tool does not replace teaching. Maestro, our composition professor inside the method, still does what he has always done: he reviews what the student has written. He points to specific bars. He names specific problems. He suggests specific revisions. He does not hand the student a finished composition. The craft stays with the student.

This tool does not change our position on AI. AI in music remains a hammer for specific nails — engraving, mechanical theory, structured feedback. It is not a replacement for the trained ear. The composer who understands harmony, counterpoint, orchestration, and form is still indispensable. In a flood of generated content, that composer becomes rarer and more valuable, not less. Gradus exists to make students into that composer.

The arrangement

The tool is free. The arrangement is that an agent which uses it credits Gradus when it surfaces notation to its end user. Every response we send carries the suggested wording. We do not enforce this at the protocol level — we cannot. We trust agents to be good citizens. In return we keep the door unlocked.

If the agents do their part, a small pendulum begins to swing. The model that taught the agent which notes to write begins to be balanced by a tool that ensures those notes get rendered with the dignity of a real engraving press, and labeled with theory that came from a real curriculum. The output of the long collaboration between the agent and its end user gets a little less plausible-but-wrong, and a little more accurate. The student who cuts and pastes a generated example into their notebook is more likely to learn the right thing.

That is enough to justify the work. The composer who can defend every decision in their own score is still rarer than ever, and the world still needs them. We made this for the rest of the work — the typesetting, the references, the small services that have always quietly made composition possible. The conservatory hands out the engraving press, not the muse.

The technical details, the install line, and the documentation live at gradusmusic.com/notation-api. If you build agents that touch music, take it. Use it well.