Cochlea

What richer-richard/cochlea MCP server does

The richer-richard/cochlea MCP server is designed for agent-driven audio workflows that need measurable results instead of subjective listening. It can turn a score into deterministic PCM, inspect an existing audio file, create visual spectrograms, and check whether a rendered result meets assertions embedded in the score.

Scores can be authored as RON data or through the Rust builder API. A score may define sample rate, timing resolution, tempo, tracks, instruments, notes, and parameter automation. Rendered output can be written as WAV audio, with optional stems and verification.

For audio review, the tool reports values such as integrated loudness, true peak, loudness range, onset timing, estimated pitch, melody events, key, tempo candidates, rhythm alignment, stereo width and correlation, structure, silence, and clipping. Reports can be emitted as JSON, while digest output provides a compact text summary suitable for a limited context window.

How it works

A typical workflow is compose, render, probe, inspect, and verify. Rendering converts the score to offline PCM. Probing reduces an audio file to structured measurements, and spectrogram generation creates an image for cases where numeric data is not enough. Verification evaluates conditions such as peak limits, pitch agreement, parameter monotonicity, and expected silence.

The richer-richard/cochlea MCP server supports time windows for read operations. A caller can focus on a precise section of a longer file, with frame-exact boundaries and timing reported relative to the selected region. Spectrograms can also be tiled into contact sheets, and annotations can show beats, onsets, and pitch.

The command-line examples also cover importing Standard MIDI Files into score data, transcribing audio into a score, linting scores, and comparing two renders. Render comparison can include a second analysis tier and a spectrogram of the difference.

Setup and configuration

The supplied setup instruction is:

cargo install cochlea-mcp

The material describes the engine as fully offline. No API key, hosted service, audio device, or ffmpeg installation is specified. The README shows command-line usage through the cochlea executable, including commands for rendering, probing, spectrogram generation, linting, importing, transcription, and diffing.

Tools and capabilities

The richer-richard/cochlea MCP server supports:

• Rendering score data to deterministic PCM and WAV output.
• Writing separate stems during rendering.
• Probing WAV, FLAC, MP3, and OGG input without ffmpeg.
• Producing JSON reports or compact digest summaries.
• Measuring loudness, peaks, pitch, melody, tempo, key, rhythm, stereo, structure, silence, and clipping.
• Generating spectrograms, annotated views, and tiled sheets.
• Comparing two renders and exporting a spectrogram of their difference.
• Verifying score assertions and returning failure through the command result.
• Importing MIDI and transcribing audio into score data.

Limitations and notes

The engine is headless: it does not use an audio device, and the documented workflow does not ask an agent to listen to PCM directly. Analysis results are estimates or measurements, so reports include confidence or stability fields where applicable, such as for key and tempo detection.

The richer-richard/cochlea MCP server is intended for deterministic, inspectable processing. The provided material does not document hosted access, client-specific configuration, authentication settings, or environment variables. It also does not establish that every input format is supported for rendering; the broader format list is specifically documented for probing.