Audio File MCP App

Audio File MCP App

An MCP App for playing and inspecting local audio files in an MCP host.

!Audio File MCP App running in Claude Desktop

Renders an in-conversation UI with playback, metadata, loudness, a spectrogram, and an optional annotation-lane timeline.

File metadata, loudness statistics, the current playhead position, any selected region, and the annotation lanes active at the playhead are also exposed back to the model, so follow-up tasks can refer to what the user is actually hearing and looking at.

Features

• Global loudness metrics computed to EBU R128: Integrated Loudness

(LUFS), Loudness Range (LRA), True Peak (dBTP), Sample Peak, and RMS.

• Instantaneous loudness metrics while playing or hovering: Momentary

(400 ms) and Short-Term (3 s) LUFS, plus sample-peak and RMS at the cursor position.

• Waveform colouring by spectral centroid — each waveform slice is

shaded along a tonal ramp using a low/mid/high band-energy ratio, so bright/dark regions read at a glance as bright/dark sound.

• Reassigned spectrogram with log-frequency bins (20 Hz floor) and an

Inferno colour scale; time-frequency reassignment sharpens transients and tonal partials beyond what a plain STFT shows.

• Looping region selection. Drag on the timeline to mark a region;

playback loops over it, and the region's start/end are passed back to the model alongside the file's loudness and playhead state.

• Annotation-lane timeline. Supply timeline annotations and the widget

draws a stack of thin labelled lanes between the waveform and spectrogram, aligned to the audio's own timeline. Each lane carries time spans, an optional colour, and an optional envelope that fades span opacity along the lane. Hovering a span reveals its label, and the lanes active at the playhead — plus those starting, ending, or active within a selected region — are reported back to the model.

Install

The server runs locally over stdio. Every install path below configures the same command: npx -y @counterpoint-studio/audio-file-mcp-app.

Claude Desktop

Easiest: grab the latest .mcpb from the Releases page and double-click it. Claude Desktop has Node bundled, so no extra runtime is needed.

Or add the server by hand to claude_desktop_config.json:

{
  "mcpServers": {
    "audio-file": {
      "command": "npx",
      "args": ["-y", "@counterpoint-studio/audio-file-mcp-app"]
    }
  }
}

Codex Desktop

1. Go to Settings -> MCP Servers
2. Add Server
3. Name: audiofile-mcp-app
4. Command to launch: npx
5. Arguments -y and @counterpoint-studio/audio-file-mcp-app

VS Code (Copilot, Agent mode)


Or add to .vscode/mcp.json (workspace) or your user mcp.json:

{
  "servers": {
    "audio-file": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@counterpoint-studio/audio-file-mcp-app"]
    }
  }
}

Goose

Paste this deep link into your browser (Goose Desktop must be installed; the custom URI scheme can't be a real link in a GitHub README):

goose://extension?id=audio-file&name=Audio%20File%20MCP%20App&cmd=npx&arg=-y&arg=%40counterpoint-studio%2Faudio-file-mcp-app

Or run goose configure → Add Extension → Command-line Extension and enter npx -y @counterpoint-studio/audio-file-mcp-app.

MCP Inspector

npx @modelcontextprotocol/inspector npx -y @counterpoint-studio/audio-file-mcp-app

Usage

Ask the host to show you a local audio file by its absolute path. For example:

"Show me /Users/me/Music/track.wav"

The host calls the display_audio_file tool, which renders the in-app UI with waveform, spectrogram, loudness metrics, and playback transport.

Annotation lanes

display_audio_file accepts an optional annotations object describing lanes to draw on the timeline (or an annotationsPath pointing at a JSON file with the same { "lanes": [...] } shape — handy for large payloads):

{
  "annotations": {
    "lanes": [
      {
        "label": "Delay tails",
        "color": "#e6007e",
        "spans": [{ "start": 8, "end": 24 }],
        "envelope": [
          { "time": 8, "value": 0 },
          { "time": 24, "value": 1 }
        ]
      }
    ]
  }
}

Times are in seconds on the audio's own timeline. label, color, and envelope are optional; a lane with an envelope fades its span opacity along the envelope curve, and uncoloured lanes use a light-accent fill. Overlapping spans in a lane are truncated at the next span's start so rows never overlap. The model can author annotations to point out sections, mark events, or visualise an analysis it just ran.

Client compatibility

Tested and known to work in:

• Claude Desktop — Chat, Cowork, and Code
• Codex Desktop
• Visual Studio Code
• Goose
• MCP Inspector

Development

pnpm install
pnpm run build:dsp   # emsdk required; see WASM-BUILD.md
pnpm run serve       # runs the server with tsx, no compile step
pnpm test

pnpm run build:dist produces the publishable layout under dist/ (dist/mcp-app.html + dist/server/).

Releasing

pnpm version <bump>          # bumps every version string (package.json, server.json,
                             # mcpb/manifest.json, src/server/app.ts) + commits + tags
pnpm publish --access public # publishes to npm, then the MCP Registry (postpublish)
pnpm run build:mcpb          # produces dist/audio-file-mcp-app-<version>.mcpb
gh release create v<version> dist/*.mcpb # attaches the bundle to a GitHub release

The MCP Registry step runs mcp-publisher publish automatically after pnpm publish. It needs a valid login — if it fails with an auth error, run mcp-publisher login github and then mcp-publisher publish by hand. mcp-publisher is a Go binary; install it via brew install mcp-publisher or per the registry quickstart.

License

ISC © Counterpoint Studio OÜ