FrameCrawler LOOPS

Track: Developer Tools

Why I built it

I was using Codex to build a cinematic water-park scene in Blender. Asking for a change was easy. Trusting the result was not.

Did it edit the right object? Was the liquid actually simulated? Is the scene still editable? Can I undo the change without damaging everything around it?

FrameCrawler makes that handoff visible.

What it does

FrameCrawler connects ChatGPT to Blender through MCP.

Instead of giving an agent open-ended control over a scene, ChatGPT creates a small typed job. Blender shows the requested boundary and waits for approval. After the change, it returns a receipt with the approval time, protected-scene boundary, revert instruction and proof artifacts.

The companion view shows a simple sequence: Requested → Changed → Proven.

Judges can inspect before-and-after renders, a portable GLB, the editable Blender scene and a short liquid preview rendered from the real Mantaflow cache.

The current proof uses a water-park scene with a 48-frame liquid bake, six moving effectors, 192 cache files and real collision. It is not a web animation pretending to be Blender output.

What made it difficult

The main problem was deciding what counts as proof.

A checksum can show that a file changed, but not that the result looks right. A screenshot can look convincing while hiding a broken or flattened scene. A server response can pass even when a human has not reviewed the creative result.

FrameCrawler keeps those surfaces separate. Deterministic gates check the receipt, approval timing, proof bundle and revert instructions. GPT-5.6 is used only after those checks, where visual intent still needs judgment.

The builder cannot approve its own work.

How I used Codex

Codex helped reduce a sprawling Blender add-on into a bounded job, receipt and recovery protocol. It also helped package the add-on, build the companion interface, tighten the bridge security and verify the real Blender proof path.

What judges can try

Install the Blender 5.1 add-on and run the local ChatGPT companion. Request one scene change, approve it in Blender, then inspect the receipt and artifacts. The package includes the add-on, companion, local MCP endpoint, proof scene and exact setup steps.

Current status

The local approval, proof and revert loop works. The remaining external gates are the public hosted OAuth sandbox and the final clean-profile human walkthrough.

Built With

• blender-5.1
• chatgpt-apps-sdk
• codex
• model-context-protocol
• node.js
• openai-gpt-5.6
• three.js
• typescript