AI agents can create scenes with Blender on macOS

Summary

In a brief 'TIL' (Today I Learned) post, Simon Willison described how he used ChatGPT Codex on macOS to gain full control over the Blender desktop application via its Python API. He iteratively improved a scene of a pelican riding a bicycle using a few follow-up prompts. Willison notes that modern frontier models can now generate editable .blend files, render static images, and even create short films by chaining rendered frames via ffmpeg.

Why it matters

This signals a practical convergence: LLM-based coding agents are no longer limited to text or 2D code; they can drive a full 3D graphics stack on a local machine. This lowers the barrier for non-programmers to create 3D content. Since the workflow requires only a standard Blender installation and natural language prompts, it offers a nearly zero-setup path for AI-assisted 3D production. For developers and content creators, this points to a new category of 'agentic' creative tools that sit between prompt-to-image models and manual 3D authoring.

Highlights

• Setup is extremely simple: install Blender from blender.org, then ask the agent to invoke the /Applications/Blender path via the Python API.
• Models can generate editable .blend files, render static images, and convert frame sequences into video using ffmpeg.
• The iteration loop works conversationally — Willison refined the scene with prompts like 'make it shittier' and 'do much better.'

Critical take

The post has weak evidentiary weight as it is a single anecdote lacking code, failure cases, or comparisons with alternatives like direct Blender scripting. The use of the 'gpt-6-astra' tag without benchmarks or reproducible prompt chains makes it difficult to distinguish genuine capability from cherry-picked successful scenarios.