Claude takes the wheel inside Blender

What it does

BlenderMCP is a two-part bridge: a Blender addon that opens a local socket server, and a Python MCP server that translates Claude’s requests into Blender commands. It lets Claude inspect your scene, spawn objects, tweak materials, adjust lighting, and even run arbitrary Python inside Blender’s environment. Recent versions added viewport screenshots so Claude can see what it is working on, plus integrations to pull assets from Poly Haven, search and download Sketchfab models, generate objects via Hyper3D Rodin, and support for Hunyuan3D.

The interesting bit

The project turns Blender into an execution engine commanded by an LLM. The clever part is the feedback loop: Claude does not just push commands blindly; it can read the scene graph back and, since v1.5.5, look at viewport screenshots to correct its own mistakes. That closes the loop between a chat interface and a complex 3D workspace.

Key highlights

• Claude can create, delete, and modify shapes, apply materials, and adjust lighting via natural language.
• Integrates with Poly Haven, Sketchfab, Hyper3D Rodin, and Hunyuan3D for asset ingestion and model generation.
• Exposes an execute_blender_code tool that runs arbitrary Python inside Blender, giving the LLM full scripting access.
• Plugs into Claude Desktop, Cursor, and VS Code through standard MCP configuration.
• Supports running the MCP server on a remote host.

Caveats

• The execute_blender_code tool runs arbitrary Python inside Blender; the README explicitly warns this is dangerous and advises saving work first.
• Poly Haven integration is described as “sometimes erratic,” and complex requests may need to be broken into smaller steps.
• Telemetry is enabled by default and must be unchecked in the addon preferences or disabled via an environment variable if you want it fully off.

Verdict

Worth a look if you want to prototype 3D scenes conversationally or automate tedious Blender workflows through an LLM. Skip it if you need deterministic, production-grade pipeline tools rather than an AI-assisted prototyping layer.