Point it at a folder of music. It finds the macOS-style duplicates (the trail of “Song (1)”, “Song (2)” files), reads a vibe profile out of each one with librosa, and asks a local Llama 3.1 model to write evocative replacements, using the names of your originals and the duplicate's own sonic character as context.
You review, edit, regenerate and approve before anything is renamed, and every session is reversible: revert a whole run or a single file from the history.
The model runs on-device through MLX. No daemon, no server, no API key.
What it does
- Detects macOS-style (1)/(2) duplicate chains
- Vibe profile per track: BPM, key, energy, brightness, mood
- Five candidate names per duplicate from a local LLM
- Per-track and batch-wide steering prompts
- Inline waveform and playback for every candidate
- Full rename history: revert a session or a single file
- Normalize mode enforces a naming convention with no AI at all