Stem separation from an AI agent, over MCP

Published October 2026

Doseedo's stem separation is one MCP tool. Connect the server and an agent — Claude, ChatGPT, Cursor, Gemini CLI, Codex — splits a mix into vocals, drums, bass, guitar, piano and other, or into a wider orchestral or auto-detected instrument set, with optional per-stem MIDI, a drum split into kick, snare, hats, toms and cymbals, a guitar split into players, and — if you ask — a native Logic or Ableton session built from the stems. The agent uploads the mix, waits, and hands back download links; because the links are reusable, the bass stem can go straight to the transcriber, the drums to a session, the vocal to a cover.

The tool

doseedo_separate_stems takes an audio_key (from doseedo_get_upload_url — WAV, AIFF, FLAC, MP3, M4A or Ogg) or an audio_url: any public URL, or a Doseedo result link such as a generated take or a bounce. The options that matter:

• models — 6stem (default: vocals, drums, bass, guitar, piano, other), or orchestra / detect / auto for wider instrument sets; with orchestra, instruments names the targets ("saxophone,trumpet,trombone").
• include_midi — transcribe each stem to MIDI as well (default on); stems_only turns off everything derived and is the fastest, cheapest run.
• drum_split — the drum stem into kick, snare, hi-hat, toms, ride, crash and percussion sub-stems.
• guitar_split — the guitar stem dealt into two to four players, or auto to let a probe decide.
• export — also build a logic or ableton session from the stems; or session_id to register the stems into an existing session as child tracks of the mix.

The price is flat and the quote tool states it as a rule. Real reply, captured 2026-10-15, for a three-and-a-half-minute mix with per-stem MIDI:

{
  "job": "separate_stems",
  "route": "/api/separate-stems",
  "estimate": { "generation": 20 },
  "basis": "flat 20 credits for 6stem / auto",
  "affordable": true
}

Orchestral and detected sets price differently — 10 × (1 + ceil(instruments / 8)), capped at 80 — and the server's own stems recipe puts the run at about two minutes. The result arrives as pre-signed links that work with a plain curl for 24 hours and can be handed straight to the next tool; the agent waits in 50-second holds with doseedo_wait_task in between.

What to ask

> Split ./mix.wav into stems, stems only.

> Separate this song and give me the drums as individual
  kick, snare and hat files.

> Separate the bounce, transcribe the bass stem, and show me
  the notes against the beat grid.

> Separate the mix and build an Ableton set from the stems,
  with MIDI for the piano and bass.

> In my "Bridge Demo" session, separate the reference track
  on track 1 into stems as its child tracks.

Separation is also the first stage of two bigger tools the agent can call directly: doseedo_audio_to_session (a mix into a complete DAW session — stems on named tracks, tempo map, markers, MIDI, chord guide) and doseedo_cover_song (a song re-rendered stem by stem with instruments swapped or a genre changed).

What it is, and what it is not

Separation is a job the server runs on the audio, so the mix uploads — to a presigned URL, and only for the run. Uploads for processing accept lossy and lossless formats on every plan. What comes back is sample-accurate stems as WAV, the per-stem MIDI when asked, and the sub-stems of a drum or guitar split, each a download link valid 24 hours; doseedo_check_task re-issues fresh links after that. It is metered from the account's credit pool, priced in advance by doseedo_quote; nothing about the MCP changes the price.

The web page does the same job with a browser and a drag-drop; the MCP version is for the moment the stems are the middle of something — a transcription, a session, a cover, a batch of forty files — that an agent then carries on with.

Related

• Stem separation, in the browser
• A mix into a DAW session
• Audio to MIDI over MCP
• Generate music and covers over MCP
• Bounce stems without opening your DAW