How Deepcut works.

From prompt to playlist in minutes.

Example: build me a 25-song late-night playlist like Khruangbin and Glass Beams

Deepcut searches several YouTube candidates, scores them, hydrates details when useful, and refuses weak matches instead of silently guessing.

list_my_playlists — show your playlists

resolve_tracks — preview likely matches before writing

build_playlist_from_tracks — main playlist builder

add_videos_to_playlist — add and dedupe tracks

update / reorder / remove / delete — manage existing playlists with confirmation safeguards

✓ Prefers official audio/video and artist releases

✓ Boosts artist/channel/title agreement

✓ Filters karaoke, covers, reactions, tutorials

✓ Avoids sped-up, nightcore, bass-boosted and loop edits

✓ Penalizes live/remix/instrumental unless requested

✓ Returns runners-up when confidence is too low

Private by default. New playlists default to private.

Confirm destructive actions. Delete/remove tools require explicit confirmation.

OAuth credentials stay server-side. They are not intentionally exposed in model-visible tool output.

Quota aware. Deepcut checks whether long jobs can finish before starting them.

Production uses one hosted MCP endpoint for ChatGPT, Claude, Cursor, and other compatible clients.

Local developers can also use stdio mode with node dist/index.js.