Audio Analyzer

audio_visualizer_rs

An MCP server that gives Claude the ability to hear music.

Point Claude at any audio file and it can tell you the key, tempo, dynamics, timbre, percussive character, stereo field, structural sections, and how the music evolves over time -- all from raw audio analysis, no images, no guessing, under 1% context window usage.

Compare two tracks side-by-side. Detect where the music changes structurally — intro, verse, chorus, bridge — and zoom into the moments that matter.

What is this?

LLMs can see (vision) and read (text), but they can't hear. This project bridges that gap by running real audio analysis -- the same DSP techniques used in music information retrieval research -- and returning structured numerical data that Claude can reason about.

It's an MCP server that exposes audio analysis as tools Claude can call on demand. Ask Claude to analyze a song and it will decode the audio, run spectral/harmonic/rhythm/percussive analysis, and return the results as compact text. No spectrograms, no images, no wasted tokens.

Full analysis of a 60-second track completes in under 2 seconds (including source separation). Pure Rust. No Python, no FFmpeg, no system dependencies.

Features

• Audio decoding -- mp3, wav, flac, ogg, aac via Symphonia (pure Rust)
• Spectral analysis -- centroid (brightness), bandwidth (richness), rolloff, flatness (tonality)
• Frequency band energy -- RMS energy across 7 standard producer bands (sub-bass through brilliance) for mix diagnosis
• Spectral contrast -- peak vs valley per band in dB, reveals clarity vs muddiness
• Dynamic range -- crest factor, loudness range (95th-5th percentile), peak dBFS
• LUFS loudness -- EBU R128 integrated loudness (ITU-R BS.1770-4 stereo channel summing), true peak (dBTP), loudness range (LRA), streaming platform targets. Validated to 0.0 dB of FabFilter Pro-L 2.
• Stereo field analysis -- phase correlation (mono compatibility), stereo width (mid/side ratio), L/R balance, mono compatibility score. Per-frame time-series to pinpoint where phase issues occur.
• Temporal features -- RMS energy (loudness), zero crossing rate (texture)
• Timbre -- 13 MFCCs (Mel-frequency cepstral coefficients)
• Harmonic analysis -- chromagram, key detection (Krumhansl-Schmuckler algorithm), tonnetz
• Rhythm analysis -- tempo estimation, beat tracking, onset detection, tempo stability
• Percussive character -- harmonic/percussive source separation (HPSS), attack sharpness, onset density
• Section boundary detection -- multi-feature novelty analysis detects structural changes (energy, spectral, harmonic, texture). Enables a summary→zoom workflow: get the map, then dive into interesting moments
• A/B comparison -- analyse two tracks side-by-side, get a compact diff table highlighting differences in loudness, dynamics, spectral balance, stereo field, key, and tempo
• Time-series data -- track how every feature evolves over time at selectable resolution
• Token-efficient -- downsampled output calibrated to fit comfortably in the context window

See it in action

Watch a fresh Claude analyse four tracks in one session — an original piano/brass composition, Bohemian Rhapsody's full structure mapped with section boundaries, an A/B mix comparison with production advice, and a thunderstorm field recording where it triages music metrics as "meaningless" vs "physically real." All four analyses fit in a single context window.

Installation

Claude Desktop — one-click install

Download the .mcpb bundle for your platform from GitHub Releases and open it. Claude Desktop will handle the rest — no config files, no terminal, no setup.

Claude Code — Homebrew (macOS)

brew tap JuzzyDee/tapbrew install audio-analyzerclaude mcp add --scope user audio-analyzer -- $(which audio-analyzer-mcp)

Claude Code — manual (all platforms)

Download the mcp-server binary for your platform from GitHub Releases, then:

claude mcp add --scope user audio-analyzer -- /path/to/mcp-server

Build from source

git clone https://github.com/JuzzyDee/audio-analyzer-rs.gitcd audio-analyzer-rscargo build --releaseclaude mcp add --scope user audio-analyzer -- target/release/mcp-server

Restart Claude Desktop. The audio analysis tools will be available in your conversations.

Note: This is a local MCP server using stdio transport, so it requires Claude Code or Claude Desktop. It does not work with claude.ai in the browser or mobile apps.

Usage

Important: This tool analyses files on your local machine. Give Claude the full file path (e.g., /Users/you/Music/song.mp3) -- don't try to upload or attach files to the chat. Claude will read the file directly from disk.

CLI (standalone)

cargo run --bin cli -- /path/to/song.mp3cargo run --bin cli -- compare /path/to/mix_v1.mp3 /path/to/mix_v2.mp3

MCP tools

Once configured, Claude can call these tools directly:

Example: full_analysis output

Here's what full_analysis returns for a 60-second jazz trio track:

═══ Full Audio Analysis ═══File: /music/jazz_trio.mp3Duration: 60.62 sec | Sample rate: 48000 Hz | Samples: 2909952Analysis completed in: 2.02s
── Spectral/Temporal Features ──Centroid (brightness):  2812 Hz — moderateBandwidth (richness):   3933 Hz — complexRolloff (energy focus): 6489 HzFlatness (tonality):    0.0824 — strongly tonalRMS Energy (loudness):  0.1160Zero Crossing Rate:     0.0402 — mixedMFCCs (timbre):         [-141.3, 13.7, 0.9, 7.7, -2.5, 3.1, -1.7, 2.3, -0.9, 1.0, -0.5, 0.9, 0.6]
── Frequency Band Energy ──Sub bass  (20–60 Hz):     0.007803Bass      (60–250 Hz):    0.013043Low-mid   (250–500 Hz):   0.004906Mid       (500–2k Hz):    0.002165Upper-mid (2k–4k Hz):     0.000203Presence  (4k–6k Hz):     0.000166Brilliance(6k–20k Hz):    0.000089
── Spectral Contrast (peak–valley dB) ──Sub bass  (20–60 Hz):     10.4Bass      (60–250 Hz):    20.4Low-mid   (250–500 Hz):   26.0Mid       (500–2k Hz):    30.9Upper-mid (2k–4k Hz):     19.4Presence  (4k–6k Hz):     16.6Brilliance(6k–20k Hz):    55.2
── Harmonic Content ──Estimated key: E minor (confidence: 0.538)Top pitch classes:   1. G  0.627 ███████████████   2. E  0.561 ██████████████   3. C  0.512 ████████████   4. D# 0.495 ████████████   5. F# 0.487 ████████████   6. C# 0.459 ███████████
── Rhythm ──Tempo: 84.0 BPM (confidence: 0.316)Beats detected: 69Mean tempo: 84.3 BPM | Median: 84.0 BPMStability: 0.951 (0=free, 1=locked)
── Percussive Character ──Percussive ratio:    0.277 — harmony-dominatedOnset density:       7.0/sec — very densePeak attack sharp:   1.000
── Dynamic Range ──Peak:            -0.44 dBFSCrest factor:    16.2 dB — very dynamicLoudness range:  76.4 dB — very dynamicQuiet sections:  -87.5 dBFS | Loud sections: -11.1 dBFS
── Loudness (EBU R128) ──Integrated:      -12.1 LUFSTrue peak:       0.0 dBTPLoudness range:  3.7 LUSpotify (-14):   turned DOWN 1.9 dB | Apple (-16): turned DOWN 3.9 dB | YouTube (-14): turned DOWN 1.9 dB
── Stereo Field ──Phase correlation:   0.257 avg, -0.822 min — some phase issuesPhase warnings:      20.6% of frames have negative correlationStereo width:        0.812 avg, 2.747 max — wideBalance:             -0.095 — slightly leftMono compatibility:  0.620 avg, 0.117 min — significant mono loss
── Section Boundaries ──  Working BPM:     84 (confidence: 0.32)  Boundaries:      2     0:18.3s  energy+spectral+texture (confidence: 0.85)     0:42.1s  energy+harmonic (confidence: 0.67)

When you add resolution: "medium", the output also includes a time-series table showing how every feature changes over the track's duration -- letting Claude see the intro build, the dynamic solo section, and the quiet outro.

Time-series resolution

All analysis tools accept an optional resolution parameter that controls time-series output:

You can also pass a numeric string (e.g., "20") for custom rates.

Without resolution, tools return summary statistics only (averages across the whole track). With it, you get a compact TSV table showing how features evolve over time -- centroid, RMS, dynamic range, chroma, onset strength, percussive ratio, band energy, spectral contrast, and more, all aligned to the same time axis.

The presets are calibrated for token efficiency. A 3-minute track at "medium" resolution produces roughly 180 rows of data -- enough to track musical structure without blowing up the context window.

Architecture

audio file    |    +---> load_audio()          -- Symphonia decodes to mono f32 samples    |         |    |         v    |     compute_spectrogram() -- STFT via rustfft, time-frequency matrix    |         |    |         +---> spectral.rs    -- centroid, bandwidth, rolloff, flatness, MFCCs, band energy, contrast    |         +---> temporal.rs    -- RMS energy, zero crossing rate, dynamic range    |         +---> harmonic.rs    -- chromagram, key detection, tonnetz    |         +---> rhythm.rs      -- onset detection, tempo, beat tracking    |         +---> percussive.rs  -- HPSS (source separation), attack sharpness, onset density    |         +---> sections.rs    -- section boundary detection (multi-feature novelty)    |    +---> load_audio_stereo()   -- preserves L/R channels              |              +---> stereo.rs      -- phase correlation, width, balance, mono compatibility              +---> temporal.rs    -- LUFS loudness (ITU-R BS.1770-4 stereo channel summing)    |    vdownsample.rs                   -- bin-average to target resolution, format as TSV

Two binaries share the same analysis library:

• cli (src/main.rs) -- runs all analyses and prints results
• mcp-server (src/mcp_server.rs) -- exposes tools over stdio JSON-RPC via rmcp

Key dependencies: symphonia (audio decoding), rustfft (FFT), rmcp (MCP SDK).

Changelog

See CHANGELOG.md for version history.

License

MIT License. See LICENSE for details.