Building a browser music visualizer with Goose

#TL;DR

Goose (Block’s open agent for software tasks) helped me go from “I want a local music visualizer” to a working React + Vite + MUI app with Wavesurfer.js and three render modes — faster than hand-rolling every file. The interesting engineering was not the UI chrome; it was Web Audio graph wiring (one MediaElementSource per element, AudioContext resume on user gesture, cleanup on unmount).

#What shipped

Stack: Vite, React 18, TypeScript, wavesurfer.js v7, Material UI. No backend — pure client.

The audio graph is the product:

<audio> → MediaElementSource → AnalyserNode → (Wavesurfer | canvas FFT loop)

Wavesurfer.js owns waveform rendering; Web Audio API owns frequency data for bar/circular modes. Mixing them on one AudioContext is standard, but lifecycle bugs dominate the debugging time.

#Why Goose here

Goose is strongest when the task has clear boundaries and lots of repetitive structure: component files, hook skeletons, MUI layout, icon imports. I described the target UX in plain language:

• Upload a local audio file
• Show waveform while playing
• Switch between bar spectrum and a circular FFT-style view
• Dark theme, minimal controls

Goose generated the initial component tree and hook API. I then iterated on the audio pipeline manually — agents are helpful for scaffolding, not for subtle browser audio rules.

Practical tips when pairing with Goose (or any coding agent):

1. Pin libraries upfront — “Wavesurfer v7, no legacy plugins.”
2. Ask for one vertical slice — file upload → decode → play → static waveform before fancy modes.
3. Review lifecycle code — useEffect cleanup is where visualizers leak GPU/CPU.

#Web Audio lessons (the real bugs)

#One MediaElementSource per <audio> element

The browser throws if you call createMediaElementSource twice on the same element. React strict mode and hot reload made this common. Pattern:

• Store _mediaElementSource on the media element
• Reuse if already created
• Disconnect on teardown

#AudioContext suspension

Autoplay policies suspend contexts until a user gesture. Every play button needs:

if (audioContext.state === "suspended") {
  await audioContext.resume();
}

Without this, the waveform renders but the analyser flatlines.

#Mode switching

Three modes share one analyser:

• waveform — Wavesurfer’s built-in renderer
• bars / circular — custom canvas loop reading getByteFrequencyData

Switching modes means canceling the previous requestAnimationFrame, detaching old canvases, and re-binding Wavesurfer without recreating the media element source.

#Scope boundaries

This repo is only the visualizer: upload, decode, play, three render modes. No backend, no graph database, no meeting-ingestion pipeline. That narrow scope is why Goose worked — the agent could own the component tree while I owned the audio graph.

#Takeaways

1. Agents accelerate scaffolding, not audio graph correctness — budget time for browser APIs.
2. Wavesurfer v7 + Web Audio is enough for a portfolio-grade demo without a backend.
3. Explicit cleanup in hooks matters more than visual polish for “why did my laptop fan spin up?”

#Agent tooling note (2025–2026)

Goose is Block’s open-source agent for developer workflows (Block Open Source). It is not a model — it orchestrates tools against your repo. For bounded UI tasks (component tree, MUI shell, hook stubs), that is enough. For Web Audio, treat the agent output as a starting point and verify against MDN: Autoplay guide and browser DevTools performance panel.

#Analyser configuration (bars / circular modes)

Custom canvas modes share one AnalyserNode wired after Wavesurfer’s media element. The hook pins FFT and smoothing explicitly — defaults are not always what you want for bar charts:

analyserRef.current = audioContextRef.current.createAnalyser();
analyserRef.current.fftSize = 1024;
analyserRef.current.smoothingTimeConstant = 0.8;

#The one-source-per-element rule

Browsers allow only one MediaElementSourceNode per <audio> element. The hook stores the node on the element as _mediaElementSource and reuses it when switching visual modes — recreating the source throws and leaves you with a silent analyser loop. Teardown disconnects analyser and source in cleanupAudioContext() before unmount.

#Autoplay policy in practice

AudioContext often starts suspended until a user gesture. The hook calls resume() before wiring nodes and logs failures instead of assuming play() alone is enough — the difference between “works on my machine after refresh” and “works in Safari on first click.”

#Closing thought

Visual modes are cheap; leaking AudioContext instances and fighting autoplay policy are what break demos. Budget time on teardown in useEffect, not on shader aesthetics.

#Related reading

• An Astro + Bun landing template experiment — another client-only side project with a clear deploy boundary
• Building a collaborative editor with CRDTs — real-time in the browser, but state sync instead of a playback analyser loop
• MDN: Using server-sent events — useful contrast when your “stream” is analyser frames, not LLM tokens (mobile AI posts cover that pattern)