ahujasid/mcp-for-blender

The problem

Blender has a famously steep learning curve, and even experienced users spend a lot of time on repetitive scene assembly: creating primitives, positioning objects, wiring materials, checking the viewport from several angles. You either click through the UI or write bpy scripts, and neither is fast when you just want a rough blocking. This project exposes Blender to an LLM through the Model Context Protocol so you can describe what you want and have the assistant create, move and shade objects in the live scene.

What's different

It is a bridge, not a rendering or generation engine. The MCP server talks to a Blender addon over a local socket, and the LLM manipulates the actual scene, including writing Blender Python when the high-level tools are not enough. The interesting part is the feedback loop: the assistant can request multi-angle, wireframe, X-ray and render views so it can check its own work, which is the weak point of most prompt-to-3D workflows. It also pulls in assets from Poly Haven, Sketchfab and Poly Pizza, and can call third-party 3D generation services. Compared with image-to-3D generators, you get editable geometry and a full scene rather than a single mesh; compared with copy-pasting bpy scripts from a chat window, the assistant sees Blender's state and can iterate.

Who it's for

Good for Blender users who already know what they want and want to speed up blocking, scene dressing and repetitive edits; also useful for developers exploring agent-driven DCC workflows. It is not for people who want a finished asset in one prompt: results depend heavily on the model and your ability to review geometry. Studio pipelines with strict determinism, version control and licensed asset requirements should look elsewhere, and anyone hoping to skip learning Blender will be disappointed — the tool amplifies existing knowledge.

Cost of adoption

Requires Blender 3.0+, Python 3.10+ and the uv package manager, plus a compatible MCP client such as Claude Desktop, Claude Code, Cursor, Codex or VS Code. Setup is scripted (a one-line installer or uvx mcp-for-blender setup), but GUI-launched clients often fail to find uvx on PATH, a known papercut. Only one server instance should run at a time. Asset and premium model generation can involve extra accounts or API keys. The README is thorough but the project is young, the package was recently renamed, and the free/paid split is worth noting. Check the security section before running arbitrary Python from a model.

Verdict

At roughly 32k stars within months of creation, this is a project riding the MCP wave rather than a long-established tool, though a steady monthly trending gain suggests interest is holding rather than collapsing. The core idea — giving a model eyes on the viewport — is genuinely useful for Blender work and more practical than one-shot 3D generation. Treat it as a power tool for people who already model, not an autopilot. Watch the rename and premium direction, and expect rough edges.