Top AI Product

Most 3D generators hand you an asset. WorldClaw hands you a place. Type “tropical pirate island” or “snow-mountain research facility” and Tencent’s open-source framework returns an entire environment made of separate textured meshes you can orbit, walk through, and edit. Hacker News put it on the front page at 156 points.

Agents all the way down

It’s not one model, it’s a pipeline of agents running coarse-to-fine, global-to-regional. A planning agent turns your sentence into a structured spec — regions, terrain, assets, materials, spatial relations. Then procedural global terrain. Then per-region object generation and placement. Claude Opus 4.8 is the agent model, calling task-specific skills wired to GPT-Image-2, SAM3, SAM3D and Hunyuan3D, plus executable Blender tools over MCP. A refinement loop re-renders the scene and fixes bad scales, wrong poses, and trees floating six feet above the ground.

Running it yourself

Open source on GitHub as Tencent-Hunyuan/Hunyuan3D-WorldClaw, 376 stars since the August 7 drop, paper at arXiv 2608.05248. You get the agent skills and the Blender toolchain and run the pipeline locally. Output is instance-level, game-engine-ready geometry — not a pretty video — which makes it usable for game prototyping, sim environments, and previz. That’s the real jump here: 3D generation went from making objects to making worlds.

You Might Also Like

• Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
• Gpt 5 5 Takes Back the Coding Crown From Claude Opus 4 7
• 27 Agents 109 Skills 88k Github Stars is Everything Claude Code Genius or Over Engineering
• From Claude Flow to Ruflo 22k Stars 5900 Commits and the Multi Agent Swarm Taking Over Claude Code
• Oh my Claudecode Turns Claude Code Into a 32 Agent dev Team and 11 8k Stars Agree