Gemini

The first is the same scripting loop every assistant supports: describe what you want, get a bpy script, run it in Blender's Scripting workspace. Tell it your Blender version first — API drift between versions causes more failures than model quality does.

The second is a live connection through the open-source blender-mcp server. Because MCP is a client-agnostic protocol, any MCP-capable client running Gemini can drive Blender through it — the Blender side of the setup is identical to the Claude walkthrough, only the client config changes.

Screenshot-driven critique is the workflow worth building a habit around. Render a frame, hand it over, and ask a specific question — why the silhouette reads badly, whether the light ratio is doing what you intended, which normals look inverted. Concrete, visual questions get concrete answers; “make this better” does not.

This pairs naturally with the code loop: generate geometry, screenshot it, feed the screenshot back, fix. It's the cheapest approximation of an art-director pass available.

• –No mesh generation — geometry is procedural, built from code, with the usual consequences for organic shapes.
• –Fewer Blender-specific tutorials and community threads target Gemini than Claude or GPT, so when something breaks there is less prior art to search.
• –Free-tier rate limits bite during tool-heavy agent sessions, which fire many small calls in quick succession.