Starling: an AI sheet music assistant you can check

Where is Starling, and what changed today?

It is inside the workspace, in two places. On its own page the chat and the score sit side by side: you bring a recording, a video, a song link or a score file, and each score it makes or opens appears as a panel you can play, switch to jianpu and select notes in. Beside an open score in the workspace, the same chat becomes a sidebar that knows which bars you selected and where playback is.

Until today Starling was in testing, and most accounts saw it marked as coming soon. From today it is open to every signed-in account, free ones included (checked October 17, 2026). There is nothing to install and no second account: sign in, open Starling, and any score you make or change is the one in your workspace.

Your chats are kept with your account, one per score beside a score, and all of them listed on Starling’s page, so a request from last week is still there to read.

What can you ask it to do?

Anything the workspace can do, said in words. Requests that name a place and a change work best, because the reply can then be checked against those bars.

• “Bars 5 to 8 are too hard for a beginner, make the left hand simpler.” Beside the score, it edits those bars and marks the notes it touched.
• “Move it up a whole step so I can sing it in A.” The key signature, chords and every note move together.
• “Write this out for flute”, “make an easier version”, “give me the melody only”. An arrangement is made as a new score, titled for the version, so the original stays as it was.
• “Show it as jianpu”, “add guitar TAB in drop D”, “play from bar 9”.
• “Transcribe this voice memo, it is about 84 BPM.” It checks the file, tells you which choice it will use and what that costs, then starts.
• “Export the PDF and the MusicXML.” MusicXML needs Plus or a credit pack.

It can also point back at the score: select a passage for you, or play from the bar it is talking about.

How do you check that it really changed the score?

Look at the score, not only at the sentence. Starling’s edits are made through tools that report what they wrote, and the workspace shows the result in three ways.

• The notes it changed are marked on the staff, so you can see the passage it acted on.
• Its reply lists what was written, in the score’s own terms, such as a clef that went from treble to bass in bar 4. If an edit would leave the score as it was, the tool refuses it with “Nothing changed” instead of reporting success.
• Every change is a saved version, and the reply ends with Undo, which takes back everything that one reply changed. An arrangement is a new score, so the original is untouched.

Then listen. Play the score against the recording, loop the bars in question, and compare what you hear with what you asked for. Our own guide, Can an AI assistant edit sheet music for you?, walks through the same checks with screenshots of each step.

Where does it still get things wrong?

It is an early beta, and the failure that matters most is a confident sentence about an edit that did not happen. On October 10, 2026, Starling told someone it had changed bar 4’s left hand to bass clef when the bar was still in treble clef. Three things lined up: an edit tool that answered “Applied” to a change that changed nothing, a text reading of the score that hid clefs changing inside a piece, and a model that trusted a picture of the page over the text. We fixed all three in the tools and wrote what went wrong and what changed.

Other mistakes are more ordinary. It can mishear a note in a recording it transcribes, misread a crowded page, take a vague request too literally, or stop partway through a long job. It is built to report only what its tools confirmed, but it can still describe a bar incorrectly, which is why the marked notes, the written-change list and Undo are the things to rely on.

We have not measured a hallucination rate for Starling. Our evaluation sends it fixed requests about a known score and checks each claim in the reply against the score as saved, but the before-and-after results are not published yet (checked October 17, 2026). Until they are, treat a reply as a claim to check, not a receipt.

What does it cost?

About 1–5 credits a reply, by how much it does, and never more than 10 (checked October 17, 2026). A short answer uses 1 or 2; rewriting a long passage uses more. When Starling starts a transcription or a band for you, it asks first and that uses the transcription’s own price, shown before it begins. Every account gets 100 free credits a month in 2026, and the pricing page has the current packs.

What does Starling not do?

It is an assistant for the score in front of it, not a replacement for notation software, and it inherits the workspace’s limits.

• Guitar TAB has no bends, slides, hammer-ons or chord diagrams, and keeps one tuning and capo per piece.
• Jianpu is a read-only view of the staff; edit in staff notation and switch back to print.
• A recording contributes its first six minutes, a PDF up to 12 pages, and notation playback runs to ten minutes.
• It works in your account: chats and scores stay there, and MusicXML, MIDI and Guitar Pro downloads need Plus or a credit pack.

If you want the same tools inside ChatGPT, Claude or another assistant instead, the MCP connection guide connects them to the same scores.