Make a song with ChatGPT

Published October 2026

Ask ChatGPT for a song and you get lyrics, a chord chart and a structure, often good ones. Ask it for the song itself and you hit the limit of a chat: there is no band to play those chords and no project for your DAW. Add the Doseedo MCP server to ChatGPT and the conversation can go the rest of the way: the parts written as notes, played by instrument models, mixed, and returned as a Logic Pro or Ableton Live project.

What ChatGPT does well on its own

Use it for the parts of songwriting that are words and theory: lyrics that scan, a verse–chorus–bridge plan, a chord progression for a mood, a melody written out as note names. With its code tool it can even write a .mid file. What it does not give you is a performance: there is no drummer, no bass player, and no Logic or Ableton project at the end. That is the half Doseedo provides.

Connect Doseedo to ChatGPT

Custom MCP servers are a feature of ChatGPT's paid plans. Open ChatGPT → Plugins (chatgpt.com/plugins), select ＋, then Add custom MCP server. Name it Doseedo, describe it ("music and DAW projects: render parts, mix, build Logic and Ableton sessions"), give it the URL https://api.doseedo.com/mcp, choose OAuth, and approve the sign-in. On older layouts the same option sits under Settings → Apps & Connectors → Advanced settings → Developer mode. Each step, with screenshots and the business-workspace caveats, is in the ChatGPT setup guide.

What to type

> Write a 45-second country song in F at 110 bpm for drums,
  electric bass, a strummed acoustic guitar and an acoustic lead.
  Intro, verse, two choruses, outro. Write every part as notes,
  play them with Doseedo, mix them, and give me a Logic project.

Naming the form, the key, the tempo and the band takes the guesswork out. ChatGPT then follows the server's compose-piece recipe, which it can read with doseedo_recipes: it checks doseedo_instruments for each instrument's call and range, writes the parts, and calls the tools below in order.

What comes back, tool by tool

These are the replies the server gave for exactly that song on 29 October 2026. The run was made from another MCP client, but ChatGPT calls the same tools and receives the same JSON.

doseedo_render_score      51 s, 20 credits, 4 stems
  Bass           pass  41 of 41 notes
  Rhythm Guitar  pass  74 of 78 strums
  Lead Guitar    pass  44 of 44 notes
  Drums          fail  snare: 53 of 259 hits missing

doseedo_render_midi       drums again, drum_engine "oneshot"
  82 s, 20 credits       good  272 of 275 notes heard

doseedo_mix               country-pop, 23 s, no credits
  master −12.9 LUFS, true peak −1.0 dB

doseedo_session_from_stems   Logic, 30 s, no credits
  4 audio track(s) · 4 MIDI track(s)

The fourth line is the one to notice. The drum model performs a part, and a performance can drop ghost notes. The receipt counts them, the server's reply suggests re-rendering on the sampled kit, and an agent that reads its receipts does so. The project was built only after every part passed.

Lyrics and a sung vocal

The instruments Doseedo plays from notes are instrumental, so over MCP ChatGPT writes and renders the band. For a sung song there are two routes. Over MCP, doseedo_generate_music takes the lyrics and a style and sketches a sung take from the prompt alone: one mixed recording with no notes behind it, so no MIDI and no separate parts, and often not a keeper. In the browser, the studio chat writes a whole song from a sentence and sings it by default: the song engine writes the melody, the chat fits one lyric line to each phrase, and every part, the vocal included, lands on its own track, ready to export.

Related

• Connect ChatGPT to Doseedo, step by step
• The same with Claude
• Songs straight into a DAW session
• Music generation over MCP