$ agent build --scene aurora-hero.ts && vibegl verify
Your agent writes the scene.
VibeGL proves it ships.
Vibe-coded 3D demos are easy now. Production graphics are not. Untested shader cost, no device tiers, no WebGPU fallback, no rollback story. VibeGL is the control plane between "the agent made something gorgeous" and "it runs on your users' phones."
Prompt or reference in; tested, responsive and accessible web component out.
Machine-first API: submit a bundle, receive JSON verdicts, PNG evidence, tiers, fallbacks and a versioned artifact. Start at /llms.txt.
Demos lie. Evidence doesn't.
$ vibegl verify ./aurora-hero # browser matrix · visual contracts · GPU evidence ✓ visual contract "hero-default-view" Δ 0.4% (pass ≤ 1.5%) ✓ WebGPU → WebGL2 fallback renders, Δ 2.1% (allowed 4%) ▲ frame budget @ mobile-tier 11.2ms avg, 19.8ms p95 (p95 over 16.6ms) ✓ context-loss recovery restored in 240ms ✓ reduced-motion variant static poster served ▲ memory: +38MB after 5min loop suspect: unpooled geometries (report §4) verdict: SHIP WITH TIER CAP — mobile gets tier-2 (half-res bloom off) artifact: cdn.vibegl.com/aurora-hero@1.3.0 · rollback: pinned 1.2.4
Seven stages between prompt and production.
Generation and inspection tools exist. Nobody owns the part where graphics code becomes deployable. That's the whole product.
Start with the skills. Graduate to the runner.
NOW · free — the Three.js WebGPU/TSL skill pack
The knowledge layer already exists and already ships: 30+ specialized agent skills for procedural graphics, post pipelines, oceans, planets, shadows, validation — installable into Claude Code, Codex, Cursor and any file-reading agent.
NEXT · waitlist — the hosted eval runner
Visual contracts, GPU timing evidence, leak loops and regression bundles as a service — the vibegl verify run above, hosted. Design partners get free runs of their agent-produced scenes and a tier/fallback report.
hello@vibegl.com →