TrackTag MCP Server

Let AI agents (Claude Desktop, Claude Code, Cursor, and any MCP-capable client) analyze music with TrackTag: drop an MP3 into your agent, say "analyze this track", and get BPM, key, genres, moods, instruments, energy and 35+ metadata fields back.

Runs entirely on your own computer — your agent app starts it, talks to it directly (no middleman server, so your audio and API key never leave your machine on the way to TrackTag), and stops it when you quit. Analyses are billed from your TrackTag credit balance (1 credit per track with the core model, 2 with ultra); failed analyses are refunded automatically.

Setup

First, get an API key: tracktag.me/studio → API & MCP → Create key (tt_live_…). Free test keys available.

⚡ One-click install (Claude Desktop) — recommended

Download tracktag.mcpb → double-click it → Claude Desktop opens an install window → paste your API key → done. No config files, no Node.js, nothing to edit. (One Claude quirk: after installing, flip the extension ON in Settings → Extensions.)

Manual config (Cursor, Claude Code, other MCP clients)

Add the server to your agent's MCP config:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "tracktag": {
      "command": "npx",
      "args": ["-y", "tracktag-mcp"],
      "env": { "TRACKTAG_API_KEY": "tt_live_YOUR_KEY" }
    }
  }
}

Claude Code:

claude mcp add tracktag -e TRACKTAG_API_KEY=tt_live_YOUR_KEY -- npx -y tracktag-mcp

Cursor (~/.cursor/mcp.json): same JSON block as Claude Desktop.

Tools

Docs

Full API reference: tracktag.me/developers

Can I analyze multiple tracks? (batching)

Yes — small batches work right in the chat. Drop several audio files into your assistant (or name a folder's worth of tracks) and say "analyze all of these with TrackTag" — the agent calls the analysis tool once per track and summarizes the results together. Each track bills its own credit, and the default rate limit (10 requests/minute) makes this comfortable for a handful of tracks at a time.

For real catalog work, use the tools built for it: TrackTag Studio runs 100-file batches with folders, progress tracking and XLSX/CSV/XML exports, and the TrackTag API handles high-volume pipelines with async jobs and webhooks. Same engine, same credits — pick the surface that fits the job: MCP for conversation, Studio for catalogs, API for automation.

"Can't I just drag the MP3 into ChatGPT or Claude and ask?"

Not if you want real, reliable metadata — and here's the honest breakdown.

Claude cannot hear audio, and ChatGPT doesn't musically analyze it. Drop an MP3 into a Claude chat and it sees the file's name and metadata — it never processes the actual sound. ChatGPT will at best transcribe any speech in the file; it won't tell you the key, the groove or the instrumentation from the waveform. Whatever "analysis" comes back is largely an educated guess from the title. This MCP server is what gives your assistant real ears.

Some models can accept audio — but a raw model is not an analysis product. TrackTag runs every track through a purpose-built pipeline: audio preprocessing, signal-level analysis of the actual waveform (tempo, tonality, timbre, instrumentation, vocal characteristics, dynamics), an extraction process tuned specifically for music metadata, and validation

... View full README on GitHub