@audiolabtools/mcp-server

AudioLab MCP server

Exposes the proprietary AudioLab analysis engine as Model Context Protocol tools, so an AI agent (Claude Desktop, Claude Code, or any MCP client) can analyze audio.

Audio sources accepted:

• Public URL ({url: "https://..."}): fetched and decoded server-side
• Local file path ({path: "./song.wav"}): read server-side from disk

The MCP server runs on stdio in the user's own process, so local-file access is the same threat model as a CLI the user installs themselves.

Nothing is uploaded by the caller in either case. Responses are curated JSON (per-sample waveform/spectrum arrays are dropped so the response is useful to an LLM, not overwhelming).

Two modes (pick by who runs it)

	mcp/server.mjs (local)	mcp/hosted-server.mjs (hosted)
Runs	the proprietary engine locally (bundled ffmpeg)	HTTP calls to the hosted API audiolab.tools/v1/*
Audio source	{url} or {path} (on-device, never leaves the machine)	{url} only
Needs	the repo + engine (private)	an AUDIOLAB_API_KEY
Distribute?	No (it contains the proprietary engine; keep it repo-only)	Yes (no engine code, safe to hand to Claude & other AIs)

Give hosted mode to Claude Desktop / Claude Code / Cursor / any MCP client:

{
  "mcpServers": {
    "audiolab": {
      "command": "node",
      "args": ["/abs/path/to/mcp/hosted-server.mjs"],
      "env": { "AUDIOLAB_API_KEY": "al_live_yourkey" }
    }
  }
}

Get a key from [e-mail] (self-serve opens after the first partners). Both modes expose the identical eight tools below.

Tools: eight, all single-shot

Tool	Lab	Input	Returns
analyze_loudness	MixLab	{url} or {path}	Integrated LUFS (EBU R128 / BS.1770-4), true-peak, loudness range, crest factor, stereo correlation, tonal balance, harshness/muddiness labels
check_target	MixLab	{url|path, target, lufs?, tp?}	Pass/fail vs Spotify -14 / Apple-Music -16 / YouTube -14 / TIDAL -14 / Amazon -14 / podcast -16 / EBU-broadcast -23 / ATSC-broadcast -24 (or "custom" + lufs+tp). Includes deltas, issue messages, ffmpeg loudnorm fix command
analyze_timeseries	MixLab	{url|path, waveformPoints?}	Short-term LUFS samples + downsampled waveform peaks over time (for graphing/UI)
get_spectrum	MixLab	{url} or {path}	FFT bins (freqs+mags) + 7-band energies + centroid/rolloff/flatness
analyze_voice	VoiceLab	{url} or {path}	Speech/silence ratio, speaking rate, SNR, noise floor, room echo, sibilance & clipping risk
get_speech_segments	VoiceLab	{url} or {path}	Voiced regions with start/end timestamps + per-segment RMS (for auto-trim, chapters, speaker-turn detection)
index_signal	SignalLab	{url} or {path}	Content-type guess, tags, clipping/silence regions, brightness & dynamics buckets, file facts
compare_loudness	MixLab A/B	{a: {url|path}, b: {url|path}}	Run MixLab on both, return both results + per-metric deltas with verdict

Provide exactly one of url or path per source.

Requirements

• Node ≥ 22
• ffmpeg on PATH: the engine's Node decoder shells out to it.

Run

npm run mcp            # or: node mcp/server.mjs
node mcp/server.mjs --selftest   # smoke test (4 tools), no network/ffmpeg needed

The server speaks MCP over stdio (what desktop clients spawn). Streamable HTTP for a remote, keyed server is a later transport addition; deploy, auth keys, and billing are gated steps (see docs/PDD.md §0/§8).

Add to Claude Desktop / Claude Code

{
  "mcpServers": {
    "audiolab": {
      "command": "node",
      "args": ["E:/New audiolab.dev/mcp/server.mjs"]
    }
  }
}

(Once the package ships on npm, this becomes {"command": "npx", "args": ["-y", "@audiolab/mcp-server"]}, gated on publish.)

Then ask:

• "Analyze the loudness of https://example.com/track.wav" → calls analyze_loudness with {url}
• "Run loudness on ./mix.wav" → calls analyze_loudness with {path}
• "A/B compare ./master.wav against ./reference.wav for loudness" → calls compare_loudness

Limits

• 200 MB per file (size cap)
• Any sample rate / channel count (engine preserves rate; multi-channel mixed to stereo for analysis)
• One file per call (no batch; agent loops for many)
• One-shot only (no streaming / realtime)

Fastest path to first paying partner

Deploy the hosted-API counterpart (api/v1/[...endpoint].mjs) + add one key + invoice them. No Stripe, no key DB, no metering platform required to start. Metering/credits/self-serve come later, behind the billing gate.