The hard part, the separation model, is free to borrow: Demucs and the models in Ultimate Vocal Remover split a song into vocals, drums, bass, guitar and piano on your own machine. An agent can wrap that in a browser player with a stem mixer, A-B loop, tempo and key changes and an export button in one sitting. What you give up is quality at the edges (Moises' own Hi-Fi models, rough chord and lyric detection), cloud speed and the phone and tablet apps most people practice with. Good for your own song library, not a match for the Pro tools.
Before you buildMoises Free allows 5 uploads a month of up to 5 minutes each, split into 2 or 4 stems. If you only split a handful of songs a month, you may not need to pay.
Build me a local stem-separation and practice app to replace Moises. Requirements: - Python 3.11 + FastAPI (uvicorn) with a plain HTML/JS front end and no build step. `python -m moises_local` serves http://127.0.0.1:4900; the page must also work from a phone on the same Wi-Fi when I pass --lan. - Separation: run Demucs (`demucs` package, model htdemucs_6s: vocals, drums, bass, guitar, piano, other) on an uploaded mp3, wav, flac or m4a. Pick Apple MPS, then CUDA, then CPU. One job at a time, with a progress bar. - Storage: ~/MoisesLocal/<song-id>/ holds the original, six FLAC stems and song.json (title, BPM, key, chords); sqlite3 at ~/MoisesLocal/library.db lists the songs with search by title. - Player: Web Audio API with one gain node per stem, mute, solo and volume, a waveform of the mix (`wavesurfer.js`), an A-B loop with keyboard shortcuts and a one-bar count-in click. - Tempo and key: speed 50-150% and transpose -6 to +6 semitones, changed independently, using `@soundtouchjs/audio-worklet`. All stems must stay in sync after every change: test with a 4-minute song and check the drift stays under 20 ms at 70% speed. - Analysis after separation: BPM and beat grid with `librosa`, and a chord timeline from chroma features with simple major, minor and 7th templates, shown under the waveform. Label it a guide, not a transcription. - Export: the current mix (stems on or off, speed, key) as WAV or MP3 through `pyrubberband` and ffmpeg, and all stems as a ZIP. - Single user. No accounts, no telemetry, local-first. The only download is the Demucs model on first run. - Out of scope: lyrics, accounts, native mobile apps, collaboration, voice conversion, DAW plugins. - README: install steps including ffmpeg, which device was used, how long a 4-minute song took to split on this machine, and a note that CPU-only separation is slow.
Opens your agent with the prompt filled in, you press enter. Or copy it and paste it into an empty folder.
The first prompt builds the core. Run these one at a time in the same folder to grow it into your own Moises.
Extend the stem app in this folder. Requirements: - A command `moises_local watch ~/Music/to-split` that queues every new mp3, wav, flac or m4a file and runs them one after another with Demucs. - Skip files already in the library, matched by a SHA-256 of the audio bytes, and move finished originals to ~/Music/to-split/done/. - Show the queue on the home page with each job's state: waiting, running, done, failed with the error. - Retry a failed job once; stop the whole queue if the disk has under 5 GB free. - Update the README with a launchd or systemd example that starts the watcher at login.
Improve chord detection in the stem app in this folder. Requirements: - Offer a second detector next to the chroma templates: Chordino through the `chord-extractor` package, or `madmom` if Chordino will not install. Run it on the sum of the harmonic stems, not the drums. - Show both timelines on the player and let me pick the one to keep per song. - Click a chord to edit or delete it, drag its edges to retime it, and save my edits in song.json so re-analysis never overwrites them. - Export the chords as a text chart with one line per bar and the lyrics left blank. - Update the README with how each detector fails on jazz and heavily distorted guitar.
Add note transcription to the stem app in this folder. Requirements: - A button on each of the bass, guitar and piano stems that runs `basic-pitch` and writes a .mid file next to the stem, with the song tempo from song.json in the header. - Show the result as a piano roll under the waveform, following playback, with notes below 20 ms length filtered out. - Let me download the MIDI and a quantized copy that snaps notes to the beat grid. - Be explicit in the UI that polyphonic transcription is approximate. - Update the README with the model download and its licence.
Add setlists and practice presets to the stem app in this folder. Requirements: - Per song, save named presets: stem mutes and volumes, speed, key and A-B loop points; one preset can be marked as the default that loads with the song. - Setlists: an ordered list of songs I can reorder by dragging, with next and previous buttons that load each song's default preset. - A practice log in SQLite: date, song, minutes played, highest speed reached; a simple weekly chart on the home page. - Export a setlist as a text file with title, key and BPM per line. - Update the README.
Most players open it to learn a song: split it, mute the part they play, loop a hard bar, slow it down. Moises does that in a minute on a phone with nothing to install, and Premium lifts the monthly upload cap that makes the free plan awkward for a whole setlist.
Estimate · Runs on your own machine with nothing to pay after the build; Demucs is free. Without a fast GPU the cost is waiting, not money. The build is token usage on your agent plan.
Checked 2026-10-19 against Moises's docs: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
| Plan | Monthly | Annual, per month | What you get |
|---|---|---|---|
| Free | $0 | $0 | 5 uploads a month, files up to 5 minutes; stems are vocals and instrumental (2) or vocals, drums, bass and other (4); chords, metronome and lyrics only for the first minute. |
| Premium | $5.99 | $3.33 | Unlimited uploads, files up to 20 minutes, pick individual instruments (background vocals, guitar, piano, strings), full-length chords, metronome and lyrics, key and tempo changes, WAV export, faster processing. |
| Pro | $29.99 | — | Everything in Premium plus Hi-Fi separation models, drum parts, dialogue, effects and soundtrack splits, Voice Studio, VST stems plugins and 48 kHz 24-bit WAV export. |
Free allows 5 uploads a month of up to 5 minutes each, split into 2 or 4 stems.
Yes. An AI coding agent such as Claude Code, Codex or Cursor can build a usable personal version in one session with the prompt on this page, running on your own machine.
About $5.99 a month on the Premium plan, $71.88 a year (checked 2026-10-19).
Moises' own Hi-Fi models and the Pro-only splits: drum parts, dialogue, effects and soundtrack; The iPhone, iPad and Android apps, which is where most people practice; Chord and lyric detection with full-length, polished results; Voice Studio, mastering and shared Studio sessions; DAW plugins and cloud processing that does not tie up your laptop. If any of those matter to you, keep paying.
Yes: Ultimate Vocal Remover (open source), Demucs (open source), Spleeter (open source).