Search for a track, or paste a YouTube, Discogs, Bandcamp, SoundCloud, or Spotify link, to add it to your graph. Click a node to see details, pin it, or hit Discover to explore similar tracks outward — tune how far it reaches with the settings tab at the bottom of the screen.
Switch to the Library tab to see every track you've added, laid out as one graph connected by similarity. Click a node to view it or add it back into your Discovery workspace.
Search and initial lookups go through the cosine.club API separately and its own similarity score is stored alongside ours but isn't what draws the graph. After discovery, a tracks audio is extracted, analysed and reduced to a 39-dimension audio fingerprint: spectral shape (centroid, flatness, rolloff, spread, zero-crossing rate), 13 MFCCs for timbre, tempo, and loudness dynamics, each averaged and varied across two 45-second segments. Each dimension is z-scored, then weighted by feature group — MFCCs 40%, spectral shape 30%, tempo 15%, dynamics 15% — so no group dominates just by having more dimensions. Every edge and distance in the graph is the cosine similarity between two tracks' fingerprints.