NaviGravity (Navidrome Agentic Curator) 🎧

NaviGravity (NG) is an intelligent MCP (Model Context Protocol) server that empowers an AI agent to act as a sophisticated music curator for your self-hosted Navidrome library.

Unlike simple search tools, NG implements a specific curation philosophy focused on quality, discovery, and non-destructive library management.

🧠 Core Philosophy

1. The "Bliss" Quality Gate: The AI acts as a critic. No playlist is created without passing a assess_playlist_quality check to ensure diversity and prevent artist repetition.
2. Virtual Tags (Non-Invasive): We treat your audio files as sacred read-only artifacts. Moods and custom tags are stored as "System Playlists" (e.g., System:Mood:Focus), keeping your file metadata clean.
3. Smart Discovery: Features "Magic List" algorithms to surface Forgotten Gems, Hidden Tracks, and divergent genres to break your filter bubble.

🤖 For Agents & Curators

We provide a specialized guide for the Large Language Models interacting with this toolset. It defines the "Curator Persona", "Bliss Quality Gate" protocols, and strategic patterns (e.g., The Time Machine or Semantic Exploration).

👉 Read the LLM Tool Usage Manifesto

Setup & Installation

Prerequisites:

• Python 3.10+
• A running Navidrome server and user account

1. Clone the repository:

   git clone <repository_url>
   cd navigravity
   

2. Create and activate a Python virtual environment:

   python3 -m venv .venv
   source .venv/bin/activate
   

3. Install dependencies:

   pip install .
   

4. Configuration: Copy .env.example to .env (create one if needed) and set your Navidrome credentials:

   NAVIDROME_URL=http://your-navidrome-instance:4533
   NAVIDROME_USER=your_username
   NAVIDROME_PASS=your_password
   

   Logging Configuration (Optional): By default, logs are output to stderr (visible in MCP Client logs). To save logs to a file:

   NAVIDROME_LOG_FILE=./logs/navidrome_mcp.log
   

🚀 Usage

Important: This is an MCP server. It runs strictly as a backend process for an AI Client (like Antigravity, Claude Desktop or Zed). You do NOT need to "visit" it in a browser.

For a deep dive into how strict coordination works without a UI, see MCP Architecture & Workflow.

Running via MCP Client (Recommended)

Add the following to your client's configuration (e.g., claude_desktop_config.json):

{
  "mcpServers": {
    "navidrome": {
      "command": "/path/to/navigravity/.venv/bin/python",
      "args": ["/path/to/navigravity/src/navidrome_mcp_server.py"]
    }
  }
}

🛠 Troubleshooting & Local Debugging

If you encounter ModuleNotFoundError or errors spawning the process, ensure you are pointing to the Python executable inside your virtual environment.

To test the server manually and explore tools with a UI, use the MCP Inspector:

# From the project root
npx @modelcontextprotocol/inspector .venv/bin/python src/navidrome_mcp_server.py

This will open a browser interface at http://localhost:6274 allowing you to interact with all tools.

🧰 Available Tools

The agent has access to the following tools:

• Unified Analysis:

  • analyze_library(mode): One tool to rule them all.
    • mode='composition': Genre distribution & library stats (Cold Analysis).
    • mode='pillars': Identifies canonical artists by album count.
    • mode='taste_profile': Analyzes recent/frequent/starred for user habits.
  • batch_check_library_presence: Verification tool to find gaps (Missing Music) in bulk.

• Discovery & Recommendation:

  • get_smart_candidates(mode): Statistical discovery engine.
    • Smart Selection: Automatically ranks candidates by smart_score (Neutral=3, Stars=+1, Heart=+5).
    • Modes: rediscover, hidden_gems, unheard_favorites, lowest_rated, divergent (breaks filter bubble).
  • get_similar_artists: Finds relational bridges. Automatically falls back to "Genre Peers" if canonical data is missing.
  • get_similar_songs: "Radio Mode" for finding sonically resonant tracks.
  • get_genres / explore_genre: Deep dive into specific genres.
  • get_genre_tracks: Fetches random tracks from a genre.
  • search_music_enriched(query, artist?, album?): Robust metadata-rich search with multi-strategy fallback. Resolves compound queries (e.g., "Daft Punk Discovery", "Miles Davis Bitches Brew") via album expansion, unicode normalization, and client-side post-filtering. Optional artist and album parameters for precise filtering.

• Curation & Management:

  • manage_playlist(name, operation, track_ids):
    • Create/Replace customized playlists.
    • Mood Convention: Use NG:Mood:{MoodName} (e.g., NG:Mood:Focus) to create virtual mood tags.
  • assess_playlist_quality: The "Bliss" check logic. Includes automatic ID Sanitization (stripping quotes/backticks).

---

📜 License & Contributing

This project is open-source under the MIT License.

Want to help? Check out CONTRIBUTING.md for our Beta Testing Guide, Developer Instructions, and Social Contract.

[!NOTE] This project depends on libraries like py-sonic which are licensed under GPLv3. While our code is MIT, bundling it with GPL dependencies may affect the licensing of distributed binaries.

---

Built with ❤️ for the self-hosted music community.