navidrome-mcp

navidrome-mcp

An MCP server for Navidrome that builds playlists whichactually match a described mood — grounded in what you own, what you have listened to,and how you yourself have labelled your music.

Why this exists

Navidrome's API alone cannot do this, for three concrete reasons:

1. Its REST filtering is exact-match only. year=1997 works; year=1990-1999silently returns nothing. There are no ranges and no AND/OR, so a query like"90s rock I haven't played in a year, max 2 per artist" is not expressible.
2. It keeps no listen history. Only a playCount and a last playDate. There isno way to ask "what do I put on at 7am on a Tuesday".
3. Nothing in the library says how a track feels. Measured on a real 9,311-tracklibrary: 32 genres with Rock alone covering 47%, BPM present on 24 tracks (0.3%),ReplayGain on 222 (2%), MusicBrainz recording IDs on 222 (2%). That last number alsorules out AcousticBrainz as a primary mood source — its audio-derived mood models arekeyed by MBID, and resolving the rest via ISRC → MusicBrainz (rate-limited to 1 req/s)would still only reach ~23% of the library.

So this server maintains its own index and joins three sources:

It also treats your existing curated playlists as your mood vocabulary. If you haveplaylists called golden hour, cranked and slow shreds, those words already meansomething specific in your library — far more than a generic genre ever will.

The mood pass

Those playlists cover maybe 40% of a library, so on their own they can only answer moodquestions about tracks you already filed. A one-time enrichment pass closes that gap: everytrack gets energy, valence, intensity, acoustic-vs-electronic, free-form descriptors, thecurated vibe it reads as, and the times of day it fits.

The inference is grounded in your own playlists — they're a hand-labelled training set inyour own words — so unfiled tracks land in the same space as the ones you filed. Resultsare cached permanently, so this runs once and afterwards only picks up new music.

Cost stays small through three things: batching ~40 tracks per request, prompt-caching thelarge identical taxonomy prefix (the first batch runs alone, so the rest hit a warm cacheinstead of all missing it concurrently), and running with thinking disabled — this isclassification, not reasoning. Set ANTHROPIC_API_KEY to enable it and MOOD_MODEL totrade quality for spend; without a key everything else still works and only the moodfilters go dark.

Design notes

• In-memory index, no database. A full library pull is ~20s for ~10k tracks; oncelocal, every compound query is a single array pass. A JSON snapshot on disk makesrestarts instant. Deliberately no native dependency, so it installs undernpm ci --ignore-scripts on Node 20.
• The index is a cache, never a source of truth. If the snapshot is missing orunreadable the server just re-syncs.
• Navidrome's own compound engine is still exposed via create_smart_playlist, forstanding playlists that should keep re-evaluating server-side. Note those rules canonly see Navidrome's own fields — not ListenBrainz listens or Last.fm tags.

Tools

Prompt: daylist — generates a Spotify-style daylist for the current hour.

Configuration

Licence

MIT