MCP Server Guide

Blender MCP Server: Let an AI Agent Drive Your 3D Scene

A community project pairs a Blender addon with an MCP server, so an agent can build and edit scenes, assign materials, fetch assets, and run Python inside Blender. It is genuinely powerful and it hands a model a Python interpreter — plan for both.

The short answer

There is no official Blender Foundation MCP server. The widely used implementation is the third-party open-source project ahujasid/blender-mcp (MIT licensed, roughly 27.5K GitHub stars, actively maintained). It works well, but it is community software driving your local Blender install, not a vendor-supported integration.

What is the Blender MCP server?

Blender MCP has two halves that people often confuse. The first is a Blender addon that runs a small socket server inside Blender itself. The second is the MCP server process, which speaks the Model Context Protocol to an AI client and forwards instructions to that socket. Blender has to be open and the addon running for anything to work — this is not a headless cloud service.

Once connected, the agent can inspect the current scene, create, modify and delete objects, set up and assign materials, and adjust the camera and lighting. It can also pull ready-made assets from Poly Haven, Sketchfab and Poly Pizza, and generate new 3D models through AI generators such as Hyper3D Rodin and Hunyuan3D — so "make me a low-poly desk lamp and put it on the table" is a realistic single instruction rather than a demo-only one.

The capability that makes it powerful is also the one that needs the most thought: the server exposes a tool that runs arbitrary Python inside Blender. Blender’s Python API can reach the filesystem and the network, so that tool is effectively code execution on the machine running Blender, with your user’s permissions.

What you should know before connecting it

Third-party, not official

Maintained by Siddharth Ahuja at github.com/ahujasid/blender-mcp under the MIT licence. The Blender Foundation does not publish an MCP server, so there is no vendor support path if something breaks.

Two components, both local

A Blender addon exposes a socket server inside Blender; a separate MCP server process bridges that socket to your AI client. Blender must be running with the addon enabled.

Scene control tools

Create, modify and delete objects; get scene and object information; create and apply materials; adjust camera and lighting. The agent reads the scene state back, so it can correct its own work.

Asset library access

Pulls models, textures and HDRIs from Poly Haven, plus Sketchfab and Poly Pizza search and import, and AI 3D generation via Hyper3D Rodin and Hunyuan3D.

Arbitrary Python execution

A tool runs Python inside Blender. Blender’s bpy API can touch files and the network, so treat this as full code execution on the host, not as a sandboxed scripting toy.

Actively maintained

The repository was last pushed within days of this writing, so it tracks current Blender versions — but as community software, breakage after a Blender release is your problem to work around.

Setup plan

Run it on a machine you can afford to break — Because the Python tool is real code execution, use a workstation or VM where a bad script is an inconvenience rather than a disaster. Do not point it at a machine holding production credentials.
Install the Blender addon — Download the addon from the ahujasid/blender-mcp repository and install it through Blender’s Preferences, Add-ons, Install from Disk. Enable it, then start its socket server from the addon panel in the 3D viewport sidebar.
Add the MCP server to your client — Most MCP clients accept a JSON entry that launches the bridge, in the shape {"mcpServers":{"blender":{"command":"uvx","args":["blender-mcp"]}}}. Check the repository README for the current command, since packaging details change.
Confirm Blender is listening first — The commonest failure is a client that connects to nothing because the addon panel’s server was never started, or Blender was closed. Start Blender, start the addon server, then start the agent.
Begin with read-only scene queries — Ask the agent to describe the current scene before letting it build anything. This proves the socket is live and shows you how it perceives your file.
Save and version your .blend files — The agent edits the live scene and there is no undo boundary around a tool call. Work on a copy, commit checkpoints, and never let it operate on the only version of a finished project.

Useful agent workflows

Blocking out a scene from a description

Turn a written brief into a rough 3D layout — geometry, camera placement and lighting — that a human then refines by hand. The tedious first pass is where it saves the most time.

Asset sourcing without leaving Blender

Ask for a specific prop or HDRI and have the agent search Poly Haven or Sketchfab, import it, and place it in the scene, rather than doing three browser round trips yourself.

Bulk material and naming cleanup

Have the agent walk a messy scene, rename objects to a convention, and reassign materials consistently — repetitive work that is slow by hand and easy to specify in a sentence.

Procedural variation

Generate many variants of a scene or object by describing the axis of variation, using the Python tool to script the permutations.

Learning the bpy API

Ask the agent to perform an operation and then show the Python it used, as a way into scripting Blender yourself.

Security checklist

Treat the Python execution tool as full code execution on the host — the bpy API can reach files and the network.
Run it on a dedicated workstation or VM, never on a machine holding production credentials or customer data.
Work on copies of .blend files and commit checkpoints; a tool call has no undo boundary around it.
Review what the agent intends to run before approving Python execution, rather than enabling blanket auto-approval.
Treat imported asset metadata as untrusted input — text that arrives from an asset search can attempt prompt injection against the agent reading it.
Keep the socket server bound to localhost; do not expose the Blender addon’s port to your network.
Remember this is community software with no vendor support, so pin a version you have tested rather than tracking main blindly.

Treat every MCP tool as an API capability, not as a harmless chat feature. Start read-only, test in a non-production account, and require human approval for changes.

Frequently asked questions

Is there an official Blender MCP server?

No. The Blender Foundation does not publish one. The widely used implementation is the third-party open-source project ahujasid/blender-mcp, MIT licensed and actively maintained, with roughly 27.5K GitHub stars.

What can an agent actually do in Blender over MCP?

Inspect the scene, create, modify and delete objects, create and apply materials, adjust camera and lighting, import assets from Poly Haven, Sketchfab and Poly Pizza, generate 3D models through Hyper3D Rodin or Hunyuan3D, and run arbitrary Python inside Blender.

Does Blender need to be open?

Yes. The addon runs a socket server inside a running Blender instance, and the MCP server bridges to it. If Blender is closed or the addon server was never started, the agent connects to nothing. It is not a headless cloud service.

Is the Python execution tool safe?

It is as safe as running any script you did not write. Blender’s Python API can reach the filesystem and the network, so it amounts to code execution on the host with your permissions. Run it on a machine you can afford to break and review what it intends to execute.

Can I use it with OpenClaw or Hermes?

Yes. Both frameworks speak MCP, so the Blender server is added the same way as any other MCP server. The constraint is architectural rather than framework-specific: Blender runs on your local machine, so the agent that drives it must be able to reach that machine.

Primary documentation

Verify current endpoints, permissions, and preview limitations in the the blender-mcp project repository before production rollout. Vendor capabilities can change faster than third-party guides.

Run an MCP-ready AI agent without server maintenance

OpenClaw Launch handles the agent host, updates, and uptime so you can focus on safe tool design.

Configure your agent