ChatGPT

The zero-install path is still the most used one: ask for a Blender Python script, paste it into the Scripting workspace, and run it. It works because bpy is heavily represented in training data — years of Stack Exchange answers, add-on source and tutorials.

1. 01Tell it your Blender version in the first message. The bpy API has breaking changes between major versions and this single sentence prevents most of the errors people blame on the model.
2. 02Ask for a script, not an explanation — “write a bpy script that…” gets runnable code; “how do I…” gets prose.
3. 03Open Scripting → New, paste, and press Run Script.
4. 04When it errors, paste the full traceback back in. The second attempt is usually correct; the failure mode is stale API names, which a traceback fixes immediately.
5. 05Ask for the script to be idempotent — deleting objects it created on a previous run — if you're going to iterate.

Pasting scripts is a one-way conversation: the model never sees whether its code worked. The fix is the same open-source blender-mcp server the rest of this ecosystem uses — it exposes Blender as an MCP server, and any MCP-capable client can drive it, including editors like Cursor and VS Code that run GPT models.

With that connection in place, the model can read the scene, run code, and take viewport screenshots to check its own work — the difference between a code generator and something that behaves like an operator. See the Claude profile for the full setup, which is identical apart from which client you point at the server.

ChatGPT reads images, and a render or viewport screenshot is a far better bug report than a description. “Why does this look wrong?” with an image attached routinely gets you concrete answers about lighting ratios, normals facing the wrong way, or scale inconsistencies that are tedious to describe in words.

• –It is not a mesh generator. Geometry comes out of code, so organic subjects — creatures, foliage, cloth — are the weak spot. Pair it with a generator for those.
• –Deprecated API calls from older Blender versions are the single most common failure. State your version up front.
• –Long scene sessions drift without an agent loop: it assumes earlier steps succeeded and builds on a scene that doesn't exist as it imagines.
• –Counted requirements (“exactly twelve steps”) are hit-or-miss — one of the things the benchmark's spiral staircase and chess set prompts are designed to expose.