Vibe Playlists

Describe a vibe, get a playlist — for Navidrome & any Subsonic server.

     

Demo GIF coming — shot list in docs/promo/repo-polish.md.

Type a scenario like "rainy evening reading a book" or "Friday night pre-drinks" and Vibe Playlists turns it into a real playlist in your music server. An LLM translates your words into human-readable audio-feature ranges (mood, energy, danceability, BPM, and genre context), matches them against an essentia analysis of your library, and creates a playlist that shows up in every Subsonic client (the Navidrome web UI, Symfonium, Feishin, play:Sub, …).

Local-first. The LLM defaults to your own Ollama. No GPU? Point it at a cheap hosted model (Claude Haiku, or any OpenAI-compatible endpoint) — one line of config.

Status: early but working (v0.1). All three parts — the companion app, the essentia analyzer, and the Navidrome plugin — are built and verified end-to-end against a live Navidrome 0.62 with an ~11k-track library. It's solo and not yet battle-tested on other setups, so issues and feedback are very welcome. The analyzer must scan your library once before playlists work.

How it works

You type a vibe ──▶ /api/vibe ──▶ LLM (Ollama by default)
                                     │  "rainy evening" → feature ranges
                                     ▼
                       query track_analysis (essentia) ──▶ score & rank tracks
                                     │
                                     ▼
                    match to your library ──▶ createPlaylist (Subsonic API)
                                     │
                                     ▼
                    playlist appears in every Subsonic client

Unlike embedding-based tools, the matching is transparent: each vibe maps to explicit, inspectable feature ranges (e.g. relaxed ≥ 0.5, energy < 0.15, bpm < 100), so you can see why a track was chosen — and tune it.

Features

• Free-text → playlist — describe any scenario; the LLM picks the audio-feature ranges.
• Pluggable LLM, local-first — Ollama by default; OpenAI-compatible and Anthropic optional.
• Works with any Subsonic server — playlists are created via the standard Subsonic API.
• Favourite-aware — starred tracks get a relevance boost.
• Genre exclusions — keep genres you don't want out of a vibe.
• No GPU required — analysis runs on CPU; a hosted LLM removes the local-model requirement.

Quick start

git clone https://github.com/craiglush/navidrome-vibe
cd navidrome-vibe
cp .env.example .env        # set NAVIDROME_* and your LLM provider
docker compose up -d
# open http://localhost:4546   (set VIBE_PORT in .env to change the port)

Requirements

• A running Navidrome (or any Subsonic-compatible server) and an account on it.
• The bundled analyzer service (set MUSIC_DIR to your library) populates the analysis DB on first run; or bring your own track_analysis SQLite DB.
• An LLM endpoint: a local Ollama (default) or a hosted OpenAI-compatible / Anthropic key.

Configuration

All configuration is via environment variables — see .env.example.

Variable	Default	Purpose
NAVIDROME_URL	http://navidrome:4533	Your Subsonic/Navidrome server
NAVIDROME_USER / NAVIDROME_PASSWORD	admin / –	Account playlists are created under
VIBE_LLM_PROVIDER	ollama	ollama | openai | anthropic
OLLAMA_URL / OLLAMA_MODEL	http://host.docker.internal:11434 / qwen2.5:7b	Local Ollama
OPENAI_BASE_URL / OPENAI_API_KEY / OPENAI_MODEL	–	Any OpenAI-compatible endpoint (Open WebUI, LM Studio, OpenAI…)
ANTHROPIC_API_KEY / ANTHROPIC_MODEL	– / claude-haiku-4-5	Hosted Claude (cheap, no GPU)
ANALYSIS_DB_PATH / VIBE_DB_PATH	/data/analysis.db / /data/vibe.db	SQLite paths
VIBE_API_TOKEN	–	If set, /api/vibe requires Authorization: Bearer <token>
VIBE_PORT	4546	Web UI / API port

Analyzer

The analyzer service scans MUSIC_DIR with essentia-tensorflow and writes per-track features into the shared analysis DB. It extracts BPM, energy, danceability, and 5 moods (happy, sad, relaxed, aggressive, party), with genre/BPM context boosts. (Arousal and valence are deferred — stored as 0.0 for schema compatibility.)

It scans on startup (SCAN_ON_START=true); to re-trigger:

# one-shot scan (recommended):
docker compose run --rm analyzer python -m analyzer scan

The analyzer's HTTP API (port 8000) is internal to the compose network and not published by default. During an active scan, the app's read queries may briefly wait (SQLite WAL, 30s timeout) — this is normal. Analysis runs on CPU (no GPU required); the first full scan of a large library takes a while, and subsequent scans skip already-analyzed files.

Navidrome plugin (optional)

A thin Navidrome plugin generates "Vibe of the Day" playlists on a schedule by calling this app's /api/vibe for each prompt you configure. It surfaces vibe playlists natively in Navidrome with no interaction required.

Requires Navidrome 0.60+ with plugins enabled (ND_PLUGINS_ENABLED=true).

1. Build it: cd plugin && make docker-build (produces vibe-playlists.ndp), or download from Releases.
2. Copy vibe-playlists.ndp into your Navidrome plugins/ directory and restart Navidrome.
3. In Settings → Plugins → Vibe Playlists, approve permissions, then set the Companion App URL, your vibe prompts (one per line), and the refresh schedule.
   • Use http://vibe:4546 only if the plugin's Navidrome and this app share a Docker network. If the app runs as a separate stack (the common case), use http://host.docker.internal:4546 (Docker Desktop) or your host's LAN address.

Heads-up: updating the .ndp (any rebuild) makes Navidrome disable the plugin and reset its config to defaults. After every plugin update, re-enable it and re-check the Companion App URL, prompts, and schedule.

Instant Vibe: to generate a single playlist on demand from inside Navidrome, type a prompt into the plugin's Instant Vibe settings field and click Save — saving re-runs the plugin and generates that one playlist (it won't repeat on restart unless you change the text). The richer "type any vibe" box with live results is in the companion app's web UI.

The interactive "type any vibe" experience lives in the companion app's web UI; the plugin is for scheduled, hands-off playlists. Generation runs asynchronously (the app returns immediately and builds the playlist in the background), so it stays within Navidrome's scheduler-callback deadline. (Instant Mix is provided by the separate navidrome-mood-plugin.)

API

POST /api/vibe
Content-Type: application/json

{ "prompt": "lazy Sunday morning with coffee", "count": 30 }

Returns the created playlist (id, name, the LLM's reasoning, the chosen feature ranges, and a track preview). Other routes: GET /healthz, GET /api/vibe/history, DELETE /api/vibe/history/<id>.

Develop / test

cd app
python -m venv .venv && .venv/bin/pip install -r requirements.txt   # Windows: .venv/Scripts/pip
PYTHONPATH=. .venv/bin/pytest -q

Roadmap

• Bundled essentia analyzer service (populates track_analysis from your library)
• Navidrome plugin — scheduled "Vibe of the Day" packs
• Optional multi-user / public-playlist sharing
• Save & re-run favourite vibes

License

GPL-3.0 — same as Navidrome.