An MCP server that lets your AI assistant analyse audio

What it adds to your assistant

The Model Context Protocol (MCP) is the open standard that lets an AI assistant call external tools. An audio MCP server is the bridge between the model and a sound file: the assistant passes a file path or a URL, the server runs the analysis and returns structured data the model can reason about.

With Tag-per-Track connected, you can ask in plain language for things an assistant could not do on its own:

• "What are the BPM and key of this track?"
• "Is this demo AI-generated?"
• "Sort the demos in this folder and tell me which three to listen to first."
• "How many monthly listeners does this artist have on Spotify?"

Watch the one-minute demo in Claude Desktop.

The tools

The server also ships three ready-made prompts: triage_demos, qualify_demo_ar and batch_demo_screening.

Install it in Claude Desktop or Cursor

1. Get a Studio API key. Sign in on Tag-per-Track, open the API & Agents tab and generate a key: it is inserted into the configuration for you. It uses the same prepaid credits as the web studio.
2. Add the server to your client's MCP configuration. In Claude Desktop this is claude_desktop_config.json; Cursor and other clients have an equivalent MCP settings file.
3. Restart the client. The Tag-per-Track tools appear in the tool list.

{
  "mcpServers": {
    "tag-per-track": {
      "command": "npx",
      "args": ["-y", "tag-per-track-mcp@latest"],
      "env": {
        "TAG_PER_TRACK_API_KEY": "tpt_live_YOUR_STUDIO_API_KEY"
      }
    }
  }
}

The server runs locally through npx, so Node.js must be installed. Full options, including wallet payment, are in the package README on npm.

Sorting a demo folder in one sentence

triage_demo_folder is the tool built for A&R work. Point it at a folder and it:

1. lists the audio files and reads artist and title from the Artist - Title file name, or from the audio tags;
2. analyses each track, with a bounded number of parallel requests;
3. looks up each artist's Spotify traction, once per artist;
4. returns a compact report ranked by score, sorted into buckets: priority, listen, pass and ai_flagged.

It handles 25 tracks per call by default and up to 50. A dryRun option lists the files and the estimated cost without analysing or charging anything, and you choose the scoring profile: discovery, signing, beatmaker or auto. More on the scoring in demo management for labels.

Two ways to pay

• Studio API key (recommended). Prepaid credits bought by card: €5 for 33 credits or €15 for 100 credits, which never expire. No crypto involved, and costs are reported in credits.
• Wallet, for autonomous agents. Per-call payment in USDC on Base through the x402 protocol: 0.15 USDC per analysis, 0.25 USDC with lyrics. A client-side spending cap, 0.50 USDC per request by default, protects the wallet.

Failed analyses are not charged, and if credits run out mid-folder the remaining files are simply not sent.

What happens to your files

• Only audio files are accepted. The server validates the format and rejects anything else, so an assistant cannot use it to send other documents from your disk.
• Large files are compressed before upload. WAV and AIFF files, and any file over 15 MB, are converted locally to AAC to cut upload time.
• Audio is not kept. Files are stored temporarily for the analysis and deleted once the report is generated. Tag-per-Track never uses your audio to train AI models.
• History is your choice. Analyses made with an API key are saved to your dashboard unless you set TAG_PER_TRACK_NO_PERSIST; wallet-paid analyses are not stored.

Two steps rely on external providers: the AI-generation check, run on a 12-second excerpt, and lyrics transcription, which is optional and sends the full track.

Not using MCP?

The same analysis is available as a REST API for your own scripts and services, and in the browser for one-off checks: BPM and key finder, AI music detector.

Frequently asked questions

What is an MCP server?

A small program that exposes tools to an AI assistant through the Model Context Protocol, an open standard. The assistant decides when to call a tool, the server does the work and returns the result. Here the tools analyse audio files.

Which AI clients does it work with?

Claude Desktop, Cursor, Windsurf and any other client that supports the Model Context Protocol. The server runs locally on your machine through npx.

Do I need a crypto wallet?

No. With a Studio API key, the server uses prepaid credits bought by card and reports costs in credits. Wallet payment in USDC is an option meant for autonomous agents.

Does the assistant actually listen to the music?

Not itself. The server sends the file to the Tag-per-Track analysis API, which measures the audio and returns data: tempo, key, genre, mood, AI verdict, scores. The assistant then reads and explains those results.

How much does an analysis cost?

One credit per track, two with lyrics transcription. Credits cost €5 for 33 or €15 for 100. Artist lookups are free, and failed analyses are not charged.

Which audio formats and sizes are supported?

MP3, WAV, FLAC, M4A, AAC, AIFF and OGG, up to 7 minutes per track. Heavy files are compressed locally before upload. A folder triage handles 25 tracks by default and 50 at most per call.

Is the server open source?

The package is published on npm under the MIT licence, and it is listed in the official MCP registry.