Audio to MIDI from an AI agent, over MCP

Published October 2026

Doseedo's transcription and its musical analysis are tools on the MCP server. Connect it and an agent transcribes a recording to notes — pitch, onset, offset, velocity, confidence — with per-instrument models behind one surface, and answers the questions around them: what key is this in, what tempo and meter, where are the downbeats, what are the chords, and is the part I just wrote harmonically sound. The analysis tools are synchronous and cost nothing, which makes them the grid an agent writes against before it renders a part or places notes in a session.

The tools

• doseedo_transcribe — audio in (an upload or any https URL, including a separated stem), notes out as JSON. Hints sharpen it: instrument ("piano", "sax", "vocals", "bass"), register, family when the instrument is unknown, polyphony — solo, poly, or section with a roster of chairs for a horn section ("trumpet,trumpet,trombone,tenor_sax"). With tempo_bpm the note times are also returned in beats. want:"f0" returns the continuous pitch contour instead of notes, for glides, vibrato and intonation.
• doseedo_analyze_song — key, bpm and a per-beat tempo map, meter, downbeats, sections and chords of a recording, in one synchronous call.
• doseedo_detect_chords — a chord chart with spans in seconds and beats.
• doseedo_beat_grid — beat and downbeat times, beats per bar, the beat map.
• doseedo_analyze_midi — the harmony of notes you wrote, before rendering or placing them.

Transcription is an async job — the server's recipe says 15 to 60 seconds, up to two credits and often none — waited on with doseedo_wait_task. The four analysis tools answer in the same call and are free of credits.

Real output, on a real stem

To make this concrete rather than described, here is doseedo_analyze_song run on a solo piano stem from our public jam set, captured 2026-10-15, trimmed to the shape (the full reply carried 421 chords and 25 tempo-map points):

{
  "status": "done",
  "kind": "analyze_song",
  "bpm": 91,
  "key": { "name": "C# major", "root": 1, "scale": "major", "confidence": 0.769 },
  "meter": "4/4",
  "beats_per_bar": 4,
  "chords": [
    { "chord": "F#maj7", "start": 21.03, "end": 24.09, "start_beat": 0, "end_beat": 5, "confidence": 0.73 },
    { "chord": "F#m7",   "start": 24.09, "end": 24.79, "start_beat": 5, "end_beat": 6, "confidence": 0.83 },
    { "chord": "Bm",     "start": 24.79, "end": 25.55, "start_beat": 6, "end_beat": 7, "confidence": 0.70 },
    …
  ],
  "tempo_map": [ { "bpm": 90.35, "time": 21.03 }, { "bpm": 85.91, "time": 53.03 }, { "bpm": 90.65, "time": 64.15 }, … ],
  "downbeats": [ 22.19, 24.79, 27.52, 30.20, … ],
  "sections": [ { "name": "Section A", "position_beats": 0 }, { "name": "Section A", "position_beats": 33.65 }, … ],
  "dynamics": { "fade_in_sec": 21.03, "fade_out_start": 219.82 },
  "truncated": [ { "path": "chords", "kept": 105, "total": 421 } ],
  "next": "The full result is kept: GET https://api.doseedo.com/api/jobs/ with your key."
}

Three things to read off it. The confidence numbers are real — 0.77 on the key, 0.6 to 0.87 per chord — and an agent should treat a 0.6 chord as a draft. The tempo map drifts (86 to 102 bpm across the take) because a solo piano playing rubato has no fixed grid, which is exactly why a part written against this recording should use the map, not the single bpm. And the piece has a long silent intro: fade_in_sec: 21 is why the first chord starts at 21 seconds.

Run on the same stem, doseedo_detect_chords — the beat-tracked chart — heard the opening as Ebm7 at 166.7 bpm (double-time, the same harmony spelled from its relative minor) with a different grid. That disagreement is honest information, not a bug: a dense or rubato recording is the hard case for both tools, and the recipe's own note says to detect chords on a separated harmonic stem of a mix, or on a clean take, for a chart you can trust. The first call to the analysis service in this capture also timed out at 45 seconds after the service had been idle; the error told the agent to wait 30 seconds and retry once, and the retry answered in a few seconds.

What to ask

> What key and tempo is ./ref.wav in, and where do the
  sections change?

> Transcribe the sax solo in ./solo.wav — tenor, monophonic —
  and show me the notes in beats at the song's tempo.

> Separate ./mix.wav, transcribe the bass stem, and put the
  notes on the Bass track of my "Bridge Demo" session.

> Here's the horn section stem: two trumpets, a trombone and
  a tenor. Give me each chair's part.

> Check the harmony of this progression I wrote before you
  render it.

The third prompt is the point of having this on the MCP: the notes do not stop at a file. The agent writes them into a session track with set_midi_notes (exact notes, played by the DAW's instrument), or renders them as audio with doseedo_render_midi on a modelled instrument, or hands them to a session build. The MIDI stays as given — the server never snaps or corrects notes behind the agent's back.

Notes, not a .mid file — and why

doseedo_transcribe returns notes as structured JSON rather than a Standard MIDI File, because that is what an agent reasons over: it can filter by confidence, quantise to the grid doseedo_beat_grid returned, split a chord into voices, or check the result against doseedo_analyze_midi before anything is committed. Where a file is the goal, the project tools write one: doseedo_separate_stems with include_midi and doseedo_audio_to_session put per-stem MIDI into a native Logic or Ableton project, and the web tool downloads a .mid directly. Audio to sheet music is the notation end of the same pipeline.

Analysis is free and synchronous; transcription is metered lightly (the recipe says up to two credits, none when a CPU model serves it); doseedo_quote prices any of them first. Calls are revocable per client from Settings → API keys.

Related

• Audio to MIDI, in the browser
• Audio to sheet music
• Stem separation over MCP
• Render the notes on a modelled instrument
• All 33 tools, what each does