Music generation from an AI agent, over MCP

Published October 2026

Three generation tools sit on the Doseedo MCP server, and they are different instruments for different jobs. doseedo_generate_music makes an original song or instrumental from a text prompt, with lyrics if you want them sung. doseedo_render_midi plays your exact notes on one of the modelled instruments — a violin, a tenor sax, an electric guitar, a piano — at your tempo, in your key, with several takes to choose from. doseedo_cover_song re-renders a song you upload, keeping its structure while swapping instruments, changing the genre or replacing the lyrics. An agent can chain them: analyse a reference, write a part that fits, render it, and drop the best take onto a session track.

The three tools

• doseedo_generate_music — prompt ("warm jazz trio, brushed drums, late-night ballad"), optional lyrics, duration_seconds, seed. bpm and time_signature are caption hints, not controls: the output can drift from them. A part that must sit on a session's grid is written as notes and rendered instead.
• doseedo_render_midi — notes in beats from bar 1 (or a midi_url, or a session track by session_id + track_id), a required instrument id from the catalog at GET /api/instruments, a required bpm, up to four takes in one call with no extra credits, a timbre_preset or a timbre_audio_url — a short recording whose sound the render takes on — a dynamics target, and voice_split for chords (one voice per render, mapped in the result). The notes are sent as written; nothing is snapped.
• doseedo_cover_song — an uploaded song plus an instruments swap map ({"piano":"electric_guitar","sax":"violin"}), a genre, replacement lyrics, a reference mix whose timbre to take on, a drum_split into isolated kit pieces, and a fidelity dial for how far to depart from the source.

All three are async jobs: one to three minutes for a render, waited on with doseedo_wait_task; the takes come back ranked best-first, as pre-signed links valid 24 hours, with the seed to re-render close to the same take. A render costs ten credits per 30-second window, and the quote tool shows its working. Real reply, captured 2026-10-15, for a 21-second upright-bass part with three takes:

{
  "job": "render_midi",
  "estimate": { "generation": 10 },
  "basis": "10 credits per 30 s window (21 s → 1 window) → 10 credits",
  "assumptions": [
    "takes (3) render in one GPU batch and cost nothing extra",
    "a chord part split into voices (voice_split) costs one batch per ~6 voices — pass voices to price it"
  ],
  "affordable": true
}

The server's recipes put the rest of the family on the same scale: generate is 10 credits per 30-second window (a four-minute song is 90), a cover is 40 and takes three to eight minutes because long songs render in 30-second windows. The account's free tier covers the first renders.

The loop an agent runs

> Analyse ./ref.wav for key and tempo. Write an eight-bar
  walking bass line that fits, check the harmony, render it
  on upright bass — three takes — and put the best one on a
  new track in my "Bridge Demo" session at bar 17.

> Generate a 60-second lo-fi hip-hop instrumental, 85 bpm
  feel, and add it to the session as a reference track.

> Cover ./demo.wav as bossa nova with the piano swapped
  for nylon guitar, and give me the stems.

> Render this melody as a horn section: split the chords
  into voices and render each one on its own instrument.

The first prompt uses five tools in order — analyze_song, analyze_midi, render_midi, wait_task, add_audio_to_session — and it is the compose-part recipe the server itself publishes through doseedo_recipes, so the agent does not have to invent the sequence. The result lands on a session track; with Dø Desktop running on a Mac with that Logic project open, it is in the arrangement within seconds.

Prompt generation versus rendering notes

This is the distinction that saves disappointment. Prompt generation is for material: a bed, a reference, a sketch of a style, a whole song with sung lyrics. It will not hold a key or a tempo you name, because the prompt is a caption, not a score. Rendering is for parts: you (or the agent) write the notes, in the key and at the tempo the session actually has, and a modelled instrument performs them — which is why it is the tool for anything that has to sit in an arrangement. A cover is the third case: the song exists, and what changes is who is playing it. The MIDI-to-audio page is the browser version of rendering and the instrument catalog lists what can play.

Generation runs on the server, so nothing about it depends on your DAW being open — it works from a phone, a terminal, a CI job. What does need the desktop bridge is only the last step, the take appearing in an open Logic project; without it, the take is on the session and in the next download.

Related

• MIDI to audio, in the browser
• The instruments
• Analyse the reference first: key, tempo, chords
• Stem separation over MCP
• Put the take on a track: editing sessions with Claude