Install once.
Then just ask.
WebGPU & WebGL profiling for your coding agent.
- Codex
- Claude Code
- Gemini CLI
- + more
Install the skill.
Run in your terminal. Choose your agent when prompted.
npx skills add Makio64/gpu-perf-agent -g
One command for all supported agents. Installed across your projects via skills.
Ask your agent.
Start a new agent session in your project.
Use gpu-perf-agent to optimize the webgpu performance
Your agent runs the profiler, makes changes, and compares the results.
Requires Node.js 22+, Git, and Chrome / Chromium. Runs on your machine.
Under the hood
Give your agent a way to measure.
The skill teaches your agent to profile locally, investigate a bottleneck, and verify its code changes with repeatable measurements.
- Capture baseline
- Agent edits code
- Capture again
- Compare
Profile your running app.
The CLI instruments your WebGPU or WebGL page in Chrome.
Optional Chrome traces add detail about browser activity.
Give the agent evidence to act on.
- In context
- A compact diagnosis keeps the agent focused.
- On disk
- The full JSON capture stays available for deeper inspection.
- In your code
- Target one suspected bottleneck and preserve the scene’s visual behavior.
Check whether the change helped.
- Compare
- Metric changes and regressions against configured thresholds.
- Quantify uncertainty
- Confidence intervals where repeated samples support them.
- Check validity
- Invalid captures, incompatible settings, and inconclusive results are flagged.
Keep the loop fast and inspectable.
- Faster iteration
- Reuse Chrome to reduce startup overhead.
- Stronger evidence
- Use longer sampling or interleaved A/B runs.
- Review & CI
- Inspect JSON or optional HTML reports; enforce performance budgets.
Frame pacing measures browser cadence; GPU duration requires timing hooks. Allocation totals estimate declared storage, not physical VRAM residency. Read the measurement reference