# Audio File MCP App

An [MCP App](https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/)
for playing and inspecting local audio files in an MCP host.

![Audio File MCP App running in Claude Desktop](https://raw.githubusercontent.com/counterpoint-studio/audio-file-mcp-app/main/docs/screenshot.png)

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](https://github.com/counterpoint-studio/audio-file-mcp-app/releases/latest)
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`:

```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)

[![Install in VS Code](https://img.shields.io/badge/Install%20in-VS%20Code-007ACC?logo=visualstudiocode&logoColor=white)](https://insiders.vscode.dev/redirect/mcp/install?name=audio-file&config=%7B%22type%22%3A%22stdio%22%2C%22command%22%3A%22npx%22%2C%22args%22%3A%5B%22-y%22%2C%22%40counterpoint-studio%2Faudio-file-mcp-app%22%5D%7D)

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

```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

```bash
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):

```json
{
  "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.

## Security model

- The server runs locally over stdio, with your user account's permissions.
- `display_audio_file` opens any file path it is given, by design, but only
  accepts files recognised as audio.
- The player streams the file through a byte-range resource
  (`audiofile-range://…`). It returns bytes only of files recognised as audio;
  anything else, including a path that doesn't exist, gets the same
  "Not a recognised audio file" error.
- `annotationsPath` reads a JSON file you name, and only uses it if it matches
  the annotation format.
- Tool errors are specific on purpose, so you can see what went wrong: they
  can reveal whether a named path exists, is a regular file, or (for
  `annotationsPath`) is valid JSON in the annotation format. They never
  include the file's contents.
- What remains: a model that can call tools can open any audio file on your
  machine. Hosts that ask for approval before tool calls are the control for
  that.

To report a vulnerability, see [SECURITY.md](SECURITY.md).

## Client compatibility

Tested and known to work in:

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

## Development

```bash
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

```bash
pnpm version <bump>          # bumps every version string (package.json, server.json,
                             # mcpb/manifest.json, src/server/app.ts) + commits + tags
git push && git push --tags
source /path/to/emsdk_env.sh # prepublishOnly rebuilds the DSP code with Emscripten
pnpm publish --access public # publishes to npm
gh workflow run publish-registry.yml -f tag=v<version> # MCP Registry, once npm shows the version
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
```

npm can hold a new version in `validating` for up to an hour before it
appears. The registry checks the npm package, so run the registry workflow
only after `npm view @counterpoint-studio/audio-file-mcp-app version` shows
the new version.

The MCP Registry is published from GitHub Actions
(`.github/workflows/publish-registry.yml`) because the interactive
`mcp-publisher login github` only grants a personal namespace, not
`io.github.counterpoint-studio/*`.

## License

ISC © Counterpoint Studio OÜ