Using Blender with Coding Agents on macOS: A Working Workflow

What this does: Lets GPT-6 Astra (via ChatGPT Codex) generate and render 3D Blender scenes through iterative natural language prompts — producing images or video sequences without writing Python manually.

Why it matters: This demonstrates that frontier models now have sufficient Blender Python API fluency to produce non-trivial 3D output. The cost floor is around $4 per multi-iteration project at API rates, putting procedural 3D generation within reach for solo operators.

Prerequisite: A macOS machine with the full Blender application installed from blender.org and access to ChatGPT Codex with GPT-6 Astra.

Steps

1. Install Blender from blender.org — the full macOS application, not a lightweight build. The agent needs access to the complete Blender Python API.

2. Open ChatGPT Codex (web or API access with GPT-6 Astra). Ensure the model has access to your local environment or the appropriate sandboxed execution context.

3. Write an initial scene prompt — be specific about what you want to render. Example: “Render a scene with a pelican on a bicycle against a sunset backdrop, with realistic lighting.” The model interprets the scene requirements and writes Blender Python scripts to construct it.

4. Let the agent execute the Python against the Blender API. The first output is a starting point — expect an initial render that has the basic composition but lacks detail.

5. Iterate with enhancement prompts — “Add a cobblestone path in the foreground” or “Improve the lighting contrast” — each prompt triggers a revised script that builds on the previous state.

6. Render the final output. For video sequences, the model can chain frames via ffmpeg — specify the sequence length and frame rate in the prompt.

7. Review cost. Simon Willison’s original workflow ran a multi-iteration project at approximately $4.24 at API rates — a useful benchmark for scoping more complex projects.

The Trap

Prompting for vague aesthetics (“make it look better”) produces marginal, unpredictable changes. Effective iteration requires specific geometric or lighting instructions: “increase the ambient occlusion intensity,” “add a point light source above and to the left,” “reduce specular highlights on the foreground objects.” The more concrete the instruction, the more targeted the Python diff the model writes.

Cited References

• Simon Willison — Using Blender with Coding Agents on macOS
• Blender Python API Documentation
• ChatGPT Codex — OpenAI
• Stack Archive: Coding Agent Components Architecture

Source: Simon Willison

Content created with AI assistance and reviewed for accuracy.