Blender MCP: Controlling Blender From Claude

Two projects answer to the name Blender MCP: the community MCP for Blender server (formerly blender-mcp) and the official Blender Lab server behind Claude’s Blender connector. What each does, how to set them up with Claude Desktop and Claude Code, and the risks of letting a model run Python in your scene.

8 min read

Blender MCP lets an AI assistant such as Claude read a Blender scene and change it by running Python through Blender’s own API. As of September 30, 2026, two projects go by that name. The one most tutorials describe is MCP for Blender, a free, MIT-licensed community project by Siddharth Ahuja that was called blender-mcp until its package was renamed mcp-for-blender. The other is the official Blender MCP server from Blender Lab, which Anthropic lists as the Blender connector in Claude. Both put a small add-on inside Blender and an MCP server beside it, and both let the model run code in your open file, so the setup is quick and the risks deserve more of your attention than the setup does.

Two projects called Blender MCP

  • MCP for Blender (community). A Blender add-on plus a Python MCP server started with uvx mcp-for-blender. It creates, edits and deletes objects, applies materials, runs arbitrary Python, and pulls in assets from Poly Haven, Sketchfab and Poly Pizza or AI-generated models from Hyper3D Rodin and Hunyuan3D. It supports Blender 3.0 or newer. Its README says plainly that it is a third-party integration and not made by Blender.
  • Blender MCP server (official). Built by Blender developers and published under Blender Lab (opens in a new tab), it needs Blender 5.1 or newer and is aimed at understanding and fixing scenes: explaining node setups, finding heavy objects, renaming data-blocks, cleaning up. It has no asset or model-generation integrations.

Pick the community server if you want Claude to build scenes and fetch assets for you. Pick the official one if you mostly want a second pair of eyes on production files and prefer software from the Blender project itself. Nothing stops you from trying both, but run one at a time against the same Blender session.

How MCP for Blender works

The MCP for Blender repository (opens in a new tab) describes two parts. The add-on (addon.py) opens a socket server inside Blender, on localhost port 9876 by default. The MCP server is a local program your client starts over stdio; it turns tool calls into JSON commands, sends them to the add-on, and passes the results back. BLENDER_HOST and BLENDER_PORT change the address, which is how you run two Blender instances side by side with two client entries on different ports. If the local and remote distinction is new to you, see local vs remote MCP servers.

Setup with Claude Desktop

  1. Install uv with its official installer (Homebrew on macOS, the shell script on Linux, the PowerShell script on Windows). The README warns against pip install uv, which may not create the uvx command your client needs.
  2. In Claude Desktop, open Settings, then Developer, then Edit Config, and add the server entry below to claude_desktop_config.json.
  3. Run uvx mcp-for-blender install-addon in a terminal. It copies the add-on into your Blender add-ons folder and keeps a backup of any file it replaces.
  4. In Blender, open Edit, then Preferences, then Add-ons, and enable Interface: MCP for Blender.
  5. In the 3D viewport, press N, open the MCP for Blender tab, and start the server. The quickstart calls the button Start MCP Server while the usage section of the same README calls it Connect to Claude, so look for either.
  6. Restart Claude Desktop and check that the Blender tools appear. Do not run the uvx command yourself; the client starts it.
claude_desktop_config.json
{
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["mcp-for-blender"]
    }
  }
}

Existing setups that still say blender-mcp keep working; the project says no config change is required, but new installs should use the new name. On Windows, if the client cannot find uvx, use its full path as the command or wrap it as cmd with the arguments /c, uvx, mcp-for-blender. General config problems are covered in the Claude Desktop MCP config guide.

Claude Code, Cursor and Codex

In Claude Code, add it at user scope so it is there in every project. The README gives a shorter form without the double dash; the Claude Code MCP documentation (opens in a new tab) recommends the -- separator for stdio servers so the server’s own flags are never read as Claude Code options. Cursor takes the same JSON as Claude Desktop in ~/.cursor/mcp.json or a project’s .cursor/mcp.json, and Codex has its own codex mcp add line. The README’s advice for all of them: run only one instance of the server, not one in Cursor and another in Claude Desktop.

Terminal
claude mcp add --scope user blender -- uvx mcp-for-blender

codex mcp add blender -- uvx mcp-for-blender

Setup for the official Blender connector

Anthropic announced the connector on April 28, 2026 in Claude for Creative Work (opens in a new tab), saying Blender’s developers built it and that, because it is built on MCP, other models can use it too. The Blender side has two steps: open the Blender Lab MCP server page, drag its install link into Blender once to add the Blender Lab extension repository, then drag it in again to install the add-on. That route also brings update notifications. The server itself is published as an .mcpb bundle for clients that accept one, or can be installed from source.

In Claude, the Claude Academy tutorial (opens in a new tab) says to open Customize, then Connectors, search for Blender and choose Add. It needs Claude Desktop, because Blender has to run on the same machine as Claude; the tutorial adds that the connector also works with other MCP clients, including Claude Code. Changes apply to your open session and reach disk only when you save in Blender.

What you can ask it to do

  • Explain and audit. “Explain what the main geometry nodes setup in this file does and add frames that document it.” Or: “List the objects with the highest polygon count relative to their size on screen.”
  • Clean up. “Fix typos in all data-block names and report which ones changed.” “Which objects use the pebbles material?”
  • Batch edits. “Make every car in the scene red and metallic.” “Point the camera at the scene and make it isometric.”
  • Assets, with the community server. “Light the scene with an overcast afternoon HDRI from Poly Haven.” Poly Haven needs no key; Sketchfab, Poly Pizza, Hyper3D and Hunyuan3D need your own keys, which you can store in the add-on preferences or environment variables. The project also offers a Premium option that generates models without your own keys.
  • Export. The community server’s export_scene tool writes GLB or FBX for use in a game engine, which is where the Unity MCP guide picks up.

Blender Lab’s own worked example is a good reminder to check the answers. Its first scene analysis only counted modifiers that affect the viewport and missed a Solidify modifier that doubled an object’s polygon count. Treat the model’s findings as leads, then confirm them in Blender.

The risks

  • It runs Python with your permissions. The community README warns that its execute_blender_code tool can run arbitrary Python and tells you to always save your work before using it. Blender Lab goes further: its server “will execute LLM generated code in Blender without any guards in place to protect your data from removal or being sent to a remote location”, and it recommends a virtual machine or a system without access to sensitive information.
  • The socket has no authentication. Anyone who can reach port 9876 can run Python inside Blender. Keep it on localhost, and use an SSH tunnel rather than pointing BLENDER_HOST at another machine.
  • Safe mode is opt-in. Setting BLENDER_MCP_SAFE_MODE=1 makes the community server check scripts first and block file access, network access, running other programs and code that keeps running afterward, while normal modeling, rendering and saving still work.
  • Telemetry. The community server records a minimal anonymous usage record by default (tool name, success, duration, versions, OS). Your prompts, code and screenshots are collected only if you opt in. DISABLE_TELEMETRY=true turns it all off.
  • Downloaded files and assets are input. A .blend file or asset from someone else can carry text that steers the model, which is indirect prompt injection. And licenses differ: Poly Haven assets are CC0, but the README says about 69% of Poly Pizza’s catalog is CC-BY and needs credit.
claude_desktop_config.json with safe mode on and telemetry off
{
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["mcp-for-blender"],
      "env": {
        "BLENDER_MCP_SAFE_MODE": "1",
        "DISABLE_TELEMETRY": "true"
      }
    }
  }
}

The wider checklist for local servers is in MCP security risks. The short version for Blender: version your .blend files, work on a copy for anything important, and keep approval prompts on in your client.

When it does not connect

  • spawn uvx ENOENT: apps started from the Dock or Start menu do not inherit your terminal’s PATH. Use the full path to uvx, then fully quit and relaunch the client.
  • Timeouts: split the request into smaller steps. Big asset downloads run on Blender’s main thread and freeze the interface until they finish, so ask for 1k or 2k textures unless an object sits close to the camera.
  • No tools at all: confirm the server is started in the Blender sidebar, then restart both Blender and the client. The README notes the first command sometimes fails and later ones work.

Turning scene audits into tasks

An audit prompt produces a list: heavy meshes, bad normals, unused materials, misnamed data-blocks. That list is more useful on a board than in a chat. With fenbs connected as a second MCP server at https://fenbs.ai/api/mcp, you can ask Claude to file each finding as a task in To Do, as a bug if something renders wrong or an enhancement if it is cleanup, with the object name and file in the note. It should search the board first so a second audit does not file the same mesh twice. If the studio has fixed conventions, such as a naming prefix for every light, record them once as a rule on the Decisions and rules page, which every connected assistant reads before it starts.

Know the limits. fenbs stores no .blend files or renders, only the tasks and their notes, and it has no due dates and no settable assignee, so a deadline or an artist’s name goes into the note as text. The board’s history does record which assistant filed or moved each task. For setting up the second server, see how to connect Claude to your project board.

Related

The same setup pattern for design files: the Figma MCP server. The engine side of a 3D pipeline: Unity MCP. Connecting assistants to a board: Claude, Claude Code and the MCP docs.

Questions people ask.

Is Blender MCP official?

There are two. MCP for Blender, formerly blender-mcp, is a community project that says it is not made by Blender. The Blender MCP server from Blender Lab is official, built by Blender developers, and is the one behind the Blender connector in Claude.

Is Blender MCP free?

Both servers are free to use. The community project is MIT-licensed and also offers an optional Premium service for AI model generation, and some asset and generation sources it connects to need your own API keys.

Which Blender version do I need?

The community MCP for Blender server supports Blender 3.0 or newer. The official Blender Lab server needs Blender 5.1 or newer.

Is it safe to let Claude control Blender?

Only with care. Both servers let the model run Python in Blender with your permissions. Save and version your files first, keep the socket on localhost, turn on safe mode in the community server, and consider a virtual machine for anything sensitive, as Blender Lab recommends.

Start with one thing.

There is nothing to set up first. Write one line and you’ve started.