MCP for Blender Pro puts a chat assistant inside Blender

A 3D artist who wants an AI assistant to arrange a scene has been stuck between a chat app and Blender, pasting instructions back and forth. MCP for Blender Pro moves the whole conversation inside Blender itself, with a Studio-style workspace, an MCP server, and the user's choice of model. The repository shows 243 stars under the MIT licence and requires Blender 4.2 or higher on Windows x64.

Writing in the project README on GitHub, Anuj Asid describes an add-on that bundles two things. The first is a chat assistant that sits in Blender and reads the current scene. The second is an MCP server, so any coding agent that speaks MCP can call the same scene tools without the Blender window being open.

What the assistant can do inside Blender

The README lists the actions the assistant exposes to a model. It inspects and analyses the scene structure. It creates and edits primitives: cubes, spheres, cameras, lights. It assigns materials and modifies properties. It duplicates and arranges objects. It searches and imports assets from Poly Haven, Sketchfab and Poly Pizza. It runs custom Python scripts through Blender's bpy API. It generates viewport screenshots so the model can see what it has just produced. And it exports the scene in several formats.

The model is the user's choice. OpenAI, Anthropic Claude, Google Gemini, OpenRouter, OpenCode Zen and Ollama all work out of the box, and the README names a custom API option for anything that speaks OpenAI Chat Completions or Anthropic Messages. A team that already has a model subscription uses that one.

The multi-agent idea, and the approval gate

The feature Asid calls out as new is a workspace that runs a team of models together instead of one. A manager model plans, worker models execute, and reviewer models check the result, with the user approving each step before it changes the scene. Blender's own undo covers the step after. Both of those are the gate against an agent that reads the scene wrong and spends twenty minutes rearranging the wrong half of it.

Treat the multi-agent part as a working feature that is young, not a finished workflow. A review loop with three agents running on three subscriptions costs three bills per task, and the README does not state how many tokens a typical scene edit uses. A buyer who runs it for a week on a real project will find out faster than a benchmark would say.

The MCP server half of the project matters for a studio that already runs Claude Code or Codex. An agent in those tools can call the Blender scene tools without the artist opening Blender, which fits a batched job the agent takes overnight. The server, the Blender chat, and the model choice sit on the same install.

The repository is MIT, which means a studio can take the Python side into its own internal build. The Windows and Blender 4.2 constraint is the firmer wall at this release: a Mac or Linux artist waits.

Source

MCP for Blender Pro, Anuj Asid on GitHub.