Let your agent guide the setup.
Sunofriend listens to separate drums, bass, keys, vocals and other song parts. It gives you editable MIDI and a clean instrumental interpretation you can hear before opening your music software.
ONE SKILL. YOUR CHOICE OF AGENT.
The plain-text Sunofriend skill tells a compatible local agent what the tool can do, what must stay local and which questions to ask before it changes anything on your computer.
Paste this into Codex with local workspace access. A normal ChatGPT conversation can explain the steps, but it cannot quietly run commands or inspect files on your Mac.
Paste this into Codex with local workspace access. A normal ChatGPT conversation can explain the steps, but it cannot quietly run commands or inspect files on your Mac.
Do not clone the repository first. These copy-ready prompts use Codex syntax. Example skills-aware agents include Codex, Claude Code and Antigravity. If $sunofriend is not recognised after turn 1, restart Codex once and send turn 2 again. In Claude Code, Antigravity or another agent, use its native skill installer or ask it to read the complete SKILL.md first.
After installing and invoking the skill, your agent should offer these same four routes one at a time.
On a supported Apple-silicon Mac, the public alpha estimates broad vocals and complementary instrumental by default. Its explicit SCNet core-four opt-in estimates vocals, drums, bass and grouped other. You listen before deciding whether any stem is useful.
Give your agent the folder location. It will check the files and, when needed, prepare supported WAV, AIFF, FLAC, M4A, MP3 or Ogg parts as one canonical WAV project before making a fresh result.
A stem is a synchronized grouped part, not necessarily one instrument. The guide explains DAW exports and independent local or cloud options when the local experimental profiles are not suitable.
Local and cloud tools have different privacy boundaries. Never upload or process music unless you have the necessary rights.Let your agent run a copyright-safe synthetic stem project through the real automatic pipeline. Then hear the result and explore the included worked MIDI pack.
The built-in demo writes only to a fresh folder you approve.The first result is deliberately useful before you learn the deeper Studio workflow.
Separate note files for the parts Sunofriend could interpret. Change notes, timing and instruments in GarageBand or another DAW.
A balanced MIDI-derived interpretation for hearing the whole idea. The source stems are level references, not audio in this mix.
The individual MIDI, combined MIDI, WAV, receipt and short start guide in one clearly labelled automatic, unreviewed bundle.
NO PERSONAL STEMS NEEDED.
Run the built-in synthetic demo to see the actual automatic MIDI/WAV/ZIP path. Then hear “Out of Place” for the wider creative goal: musical ideas carried into a new MIDI-derived interpretation.
KEEP YOUR VOICE. BUILD THE PERFORMANCE.
Sunofriend is exploring a local browser workflow for recording several attempts, comparing complete musical phrases and asking for focused pickups until your own vocal is ready to assemble. Known lyrics and a reviewed reference guide the process; your ears still make every musical choice.
See the pilot and whole-song concept →CHANGE ONE THING. KEEP THE SONG.
A bounded remix would change one instrument, role or region while keeping the motifs, harmony and groove you recognise from the source. It starts with deterministic controls and frozen models—a trained remix model is not authorised yet.
See the remix research plan →These are private research programmes, not finished products. The website cannot receive audio, and no automatic finished vocal comp or remix is available yet.
It is not an online converter. This website never receives your stems. Processing is local, and the app has only been tested on a MacBook so far. Windows and Linux remain unverified.
It is not a one-click multitrack recovery service. The public local alpha defaults to broad vocals plus complementary instrumental on a supported Apple-silicon Mac. An exact core-four MLX profile and its first PyTorch fallback are implemented but remain fail-closed after their bounded objective remediations failed. A hash-pinned SCNet-large release profile is now available as an explicit public opt-in preview for vocals, drums, bass and grouped other after its finite offline canaries and catastrophic-output listening checks passed. Separate private, unregistered six-role research combined core four with synth and guitar specialists. Its completed three-song review found all 3/3 cases useful and non-catastrophic; synth and guitar were each useful in 2/2 confirmed-present cases, with some missing content reported for both. This is positive private evidence only: resource and objective qualification remain incomplete, and there is no public six-role command, profile registration, automatic source choice or activation. Bring original or authorised parts when you need reliable keys, guitars or narrower families. See the neutral stems guide.
The working two-stem slice remains the public default. Core four is explicit opt-in. Both keep setup, rights confirmation, local inference, listening and later MIDI conversion as separate steps. Read the setup, limits, evidence and feedback guide.
It is not a perfect transcription. The WAV is a creative MIDI interpretation, not waveform reconstruction or a human-approved release master.
YOUR EXPERIENCE CAN IMPROVE THE NEXT VERSION.
Feedback from every Sunofriend user is welcome. Reports from Claude Code, Antigravity or another skills-aware agent are especially useful, as are Windows and Linux trials that could help make SKILL.md and the setup path more portable.
