MCP for Blender: model, stage and inspect scenes in Blender

Blender has a lot of menus and a lot of shortcuts, and newcomers trying to block out a simple scene often get stuck finding the right one. MCP for Blender (formerly blender-mcp) connects the Blender you have open to an agent. You describe what you want; the agent writes Blender Python to model, apply materials and light the scene, then looks at it from several angles to check its own work. It’s a community project, not affiliated with Blender, and had about 30k stars as of 2026-10-06, among the most of any MCP project.

What it does

• Code inside Blender: execute_blender_code runs Python in your live session; modeling, layout, materials, animation and rigging all go through it.
• Self-checks: look grabs the viewport, the camera, an auto-framed multi-angle sheet or an animation strip, in solid, material, rendered, wireframe or X-ray shading; get_scene_info returns a compact summary of scene objects.
• Asset libraries: search_assets and import_asset reach Poly Haven (CC0 HDRIs, textures and models), Sketchfab and Poly Pizza low-poly models.
• Generated models: generate_3d creates a model with Tripo, Hunyuan3D or Hyper3D Rodin and imports it.

Who it’s for

• Indie game developers who don’t know Blender well and want quick level blockouts or prop drafts.
• Designers making product renders who want the lights and camera set up before they polish by hand.
• Experienced Blender users who want the agent to script repetitive batch jobs such as renaming or swapping materials.

Setup

You need Blender and uv. Manually, it’s two steps: add the server to your client, then install the Blender add-on:

claude mcp add blender uvx mcp-for-blender
uvx mcp-for-blender install-addon

Then enable MCP for Blender under Edit → Preferences → Add-ons. The add-on opens a local socket server when Blender starts. The README also has a one-line installer that detects the MCP clients on your machine.

Our take

It’s the de facto standard for 3D: big user base, frequent releases, and a closed loop from writing code to looking at the result, which is what makes it more useful than servers that only send commands. The risk is real, though. execute_blender_code runs arbitrary Python, which means the agent executes code on your computer, and the project itself says to save your work first. BLENDER_MCP_SAFE_MODE=1 screens scripts and blocks file access, network calls and launching other programs. An anonymous usage record (tool name, success, versions) is sent by default; DISABLE_TELEMETRY=true turns it off. Premium AI model generation is paid. Run only one client’s server instance at a time. MIT-licensed.