webgpu-metal-debug-kit

A toolkit for debugging Chrome WebGPU applications on macOS with AI coding assistants. Combines Chrome DevTools MCP for browser-level debugging with Xcode xctrace for Metal driver-level profiling.

Why

WebGPU debugging is hard:

• No printf in shaders
• GPU timing doesn't show up in JS profilers
• When something renders wrong, you can't tell which stage broke
• Some bottlenecks are invisible from the browser (Dawn/Metal translation overhead)

This kit gives your AI assistant direct access to browser state (evaluate_script) and GPU driver traces (xctrace).

Install

Claude Code Plugin (recommended)

/plugin marketplace add penspanic/webgpu-metal-debug-kit
/plugin install webgpu-metal-debug-kit@webgpu-metal-debug-kit
/reload-plugins

As Agent Skill (manual)

git clone https://github.com/penspanic/webgpu-metal-debug-kit.git \
  ~/.claude/skills/webgpu-metal-debug-kit

Works with any agent that supports Agent Skills (Claude Code, Codex CLI, Cursor, VS Code Copilot, etc.).

How It Works

With AI (Claude Code + MCP)

The AI uses Chrome DevTools MCP to directly control the browser:

// AI runs these via evaluate_script
window.__gpu.stats()          // { fps, avgMs, maxMs, frameGapMs }
window.__gpu.setDebugMode(2)  // step heatmap
window.__gpu.textures()       // texture info

When browser-level analysis isn't enough, the AI captures Metal traces:

# AI finds GPU process and runs xctrace directly
GPU_PID=$(pgrep -f "Google Chrome.*--type=gpu-process" | head -1)
/usr/bin/xcrun xctrace record --template 'Metal System Trace' \
  --attach "$GPU_PID" --time-limit 5s --output /tmp/webgpu-metal-trace.trace

Without AI (manual)

# Start your app
python3 -m http.server 8080

# Launch Chrome with Metal debug flags
bash scripts/setup-metal-debug.sh http://localhost:8080

# Capture Metal trace
bash scripts/capture-metal-trace.sh 5

# Open in Instruments
open /tmp/webgpu-metal-trace.trace

In the Chrome console: __gpu.stats(), __gpu.setDebugMode(2), etc.

Integrate Into Your App

Add assets/webgpu-debug-helpers.js to your WebGPU app:

import { initDebugHelpers, frameStart, frameEnd, getDebugMode } from './webgpu-debug-helpers.js';

initDebugHelpers(device);

function render() {
  frameStart();
  // pass getDebugMode() to your shader uniforms
  frameEnd();
  requestAnimationFrame(render);
}

Add assets/debug-shader-snippet.wgsl to your shader for debug visualization modes.

Debug Modes

Mode	Visualization	Use Case
0	Normal	Default rendering
1	Render path	Which shader path rendered each pixel
2	Step heatmap	Raymarching performance hotspots
3	Depth	Depth buffer visualization
4	Normals	Surface normal visualization

See references/debug-modes.md for details.

Demo

bash demo/start.sh
# Opens http://localhost:8080/demo/ in Chrome

Note: file:// URLs won't work due to CORS. Use an HTTP server.

Requirements

• macOS (Metal tracing is macOS-only)
• Chrome 113+ with WebGPU
• Xcode + Metal Toolchain (xcodebuild -downloadComponent MetalToolchain if template missing)
• Chrome DevTools MCP (for AI integration)

License

MIT