"Blender MCP" usually means MCP for Blender, an open-source project by Siddharth Ahuja that lets an AI agent drive a running copy of Blender. It was called blender-mcp until November 2026, and most tutorials still use that name. With it connected, Claude Code, Codex, Cursor or Claude Desktop can create and edit objects, apply materials, look at the viewport and export files, all from a chat prompt.

This guide covers what it is, how to install it in Claude Code and Codex, and a full example: an agent builds a low-poly game prop in Blender, exports it as a GLB, the file gets compressed with glTF-Transform, and three.js loads it in a browser game. We checked every command against the project's README and the official docs on November 21, 2026. We did not run the install or the example ourselves, so treat the outputs described below as what the docs say should happen.

What Blender MCP is

MCP (Model Context Protocol) is a standard way to give an AI agent tools that act on another program. MCP for Blender has two parts:

  • A Blender add-on that opens a socket server inside Blender (port 9876 by default) and runs the commands it receives.
  • An MCP server (the Python package mcp-for-blender) that your agent starts in the background. It passes the agent's tool calls to the add-on.

The agent mostly works by writing Blender Python and running it in your open Blender file. The README lists nine tools. The ones you will see most are execute_blender_code (run Python in Blender), look (take viewport, camera or multi-angle views so the agent can check its own work), get_scene_info (a summary of the scene), and search_assets / import_asset (pull models, textures and HDRIs from Poly Haven, Sketchfab and Poly Pizza). There is also generate_3d, which makes a model with Meshy, Tripo, Hunyuan3D or Hyper3D Rodin.

It is not made by Blender. The README opens with "This is a third-party integration and not made by Blender", and the GitHub description says it is "Not affiliated with the official Blender Foundation." It is MIT-licensed and had 35,153 GitHub stars on November 21, 2026.

Blender does have its own MCP server, though. Blender Lab, part of the Blender project, publishes a smaller one at projects.blender.org/lab/blender_mcp, documented at blender.org/lab/mcp-server. Its readme calls it "deliberately small". It focuses on Blender's Python API docs, scene and file summaries, screenshots and running code. It needs Blender 5.1 or newer, is GPL-3.0, and was at v1.0.3 (September 2026). It is also listed in Claude's connector directory as "Blender", made by Blender Lab. We cover it under Alternatives.

What you need

  • Blender 3.0 or newer for MCP for Blender (the add-on declares 3.0 as its minimum).
  • Python 3.10 or newer and uv, the tool that runs the server with uvx. The one-line installer below installs uv for you.
  • An MCP-capable agent: Claude Code, Codex, Cursor, Claude Desktop, VS Code, OpenCode, Windsurf or Antigravity, per the README.
  • For the game part: Node.js (for glTF-Transform) and a three.js project.

The quick install (all clients at once)

The README's recommended path is one command that installs uv, finds the agents on your machine, lets you choose which to set up, and installs and enables the Blender add-on. It only adds a blender entry to each config file and saves a .bak copy of every file it changes.

macOS / Linux:

curl -LsSf https://www.mcp-for-blender.com/install.sh | sh

Windows (PowerShell):

irm https://www.mcp-for-blender.com/install.ps1 | iex

If you already have uv:

uvx mcp-for-blender setup

Piping a script from the internet into your shell runs it with your user's permissions. If you'd rather not, use the manual steps below, which do the same thing one piece at a time. Either way, fully quit and reopen your agent afterwards.

Install Blender MCP in Claude Code

  1. Install uv. The README gives brew install uv on macOS, curl -LsSf https://astral.sh/uv/install.sh | sh on Linux, and winget install --id=astral-sh.uv -e on Windows. It warns against pip install uv, which may not create the uvx command.
  2. Register the server with Claude Code. This is the README's command:
claude mcp add blender uvx mcp-for-blender
  1. Check it. Claude Code's MCP docs list claude mcp list (shows a status such as "✔ Connected") and claude mcp get blender, and the /mcp panel inside a session shows the same.

By default claude mcp add saves the server for the current project only. Add --scope user to use it in every project. The docs' general form puts the server command after --: claude mcp add [options] <name> -- <command> [args...]. You need the -- if you add options such as --scope or --env.

To turn off the project's anonymous usage telemetry, the README says to set the environment variable DISABLE_TELEMETRY=true for the server. In Claude Code you pass it with --env before the --.

Install Blender MCP in Codex

The Codex CLI, the Codex desktop app and the IDE extension share one config file, ~/.codex/config.toml, so one setup covers all three.

Plain MCP server. The README's command:

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

Or add it to ~/.codex/config.toml by hand:

[mcp_servers.blender]
command = "uvx"
args = ["mcp-for-blender"]

Run codex mcp list and check that blender shows as enabled. The tools appear the next time you start Codex, and /mcp in the Codex terminal UI lists active servers. OpenAI's Codex MCP docs also document an env table for environment variables, startup_timeout_sec (default 10) and tool_timeout_sec (default 60). In the desktop app, go to Settings → MCP servers → Add server, name it blender, choose STDIO and enter uvx mcp-for-blender.

Codex plugin. The README recommends this route for Codex. It adds a live Blender viewport inside the chat, thumbnail pickers for Poly Haven, Sketchfab and Poly Pizza, and @-mentions for scene objects. It isn't in the Codex plugin directory yet, so you add it from the repo:

git clone https://github.com/ahujasid/blender-mcp.git
codex plugin marketplace add ./blender-mcp/integrations/codex

Then restart Codex, install MCP for Blender from the plugins directory, and ask Codex to use it.

Cursor and others

In Cursor, open Settings → MCP and add a global server, or create .cursor/mcp.json in your project, with:

{
    "mcpServers": {
        "blender": {
            "command": "uvx",
            "args": [
                "mcp-for-blender"
            ]
        }
    }
}

On Windows the README uses "command": "cmd" with "args": ["/c", "uvx", "mcp-for-blender"]. The README also has one-click install buttons for Cursor and VS Code, and config blocks for Claude Desktop, OpenCode and Antigravity. It says to run only one copy of the server at a time, for example Cursor or Claude Desktop but not both.

Connect the Blender add-on

The quick installer does this for you. By hand:

  1. Run uvx mcp-for-blender install-addon. It copies the add-on into your Blender add-ons folder as blender_mcp.py.
  2. In Blender, open Edit → Preferences → Add-ons and enable Interface: MCP for Blender. If it doesn't show up, restart Blender, or click Install… and pick the copied file. You can also download addon.py from the repo and install that.
  3. In the 3D viewport, press N to open the sidebar and find the MCP for Blender tab. The add-on starts its server when Blender opens. If it isn't running, click Start MCP Server.

The add-on's tab is where you switch on optional features such as Poly Haven, Poly Pizza (needs a free API key) and the 3D generators.

If your agent can't connect: make sure Blender is open with the add-on running, and don't start uvx mcp-for-blender yourself in a terminal (the README's troubleshooting table says not to). Apps started from the Dock or Start menu may not find uvx and fail with spawn uvx ENOENT. Fix that by using the full path from which uvx (macOS/Linux) or where uvx (Windows) as the command. If it still fails, restart both the agent and Blender.

Worked example: a low-poly game prop, from prompt to three.js

Goal: a treasure chest for a browser game, small enough to load fast, in a format three.js reads directly. We wrote this workflow from the docs and did not run it end to end, so file sizes and exact results will vary.

1. Save your Blender file first

Open a new Blender file and save it (for example chest.blend) before the agent starts. The agent changes your file directly, and the README's own warning is to always save your work first.

2. Prompt the agent

In Claude Code or Codex, with Blender open and connected:

Use the Blender MCP tools. In the current scene, build a low-poly treasure chest
for a browser game:
- About 1 m wide, 0.6 m deep, 0.7 m tall, sitting on the ground at z = 0,
  origin at the bottom centre.
- A box body and a curved lid made from a half-cylinder with 8 segments,
  plus metal bands and a lock plate. Keep it under 1,500 triangles.
- Flat-shaded materials using Principled BSDF base colours only
  (dark wood, brass), no image textures.
- Apply all modifiers and transforms, join everything into one object named "Chest".
- Check it with the look tool from several angles and fix anything that
  overlaps or floats.
- Then select only "Chest" and export it as a GLB to ./public/models/chest.glb.
Show me each script before you run it.

A few choices in that prompt matter for games. Plain base colours export cleanly, because the glTF exporter takes base colour from the Principled BSDF node, per Blender's glTF manual. One joined object means one mesh in three.js. Putting the origin at the bottom means the chest sits on the floor when you set its position to y = 0. A triangle budget stops the agent from adding subdivision.

3. The export step

The agent writes the Blender Python. The export usually ends with a call like this one, which uses the parameters documented in Blender's export_scene API:

import bpy

bpy.ops.export_scene.gltf(
    filepath="/path/to/your-game/public/models/chest.glb",
    export_format='GLB',   # single binary file: mesh, materials and images packed together
    use_selection=True,    # only the selected "Chest" object
    export_apply=True,     # apply modifiers on export
    export_yup=True,       # glTF's +Y up convention, which three.js also uses
)

Read this script before you approve it, as with every script the agent runs. Check that the path points inside your project and that nothing deletes or overwrites files you care about.

4. Compress the GLB with glTF-Transform

glTF-Transform is a command-line tool for optimizing glTF files. Install it and run optimize. These are the commands from its README:

npm install --global @gltf-transform/cli
gltf-transform optimize input.glb output.glb --texture-compress webp

For our chest:

gltf-transform optimize public/models/chest.glb public/models/chest.opt.glb --texture-compress webp

By default optimize compresses geometry with meshopt (--compress meshopt), caps textures at 2048 px and flattens the scene graph. Our chest has no image textures, so --texture-compress webp has nothing to do here, but leave it in for props that do. To compress only the geometry, glTF-Transform also has gltf-transform meshopt input.glb output.glb --level medium.

5. Load it in three.js

A meshopt-compressed file needs the meshopt decoder. Without it, GLTFLoader throws "setMeshoptDecoder must be called before loading compressed files". The imports below are the ones three.js uses in its own examples:

import { GLTFLoader } from 'three/addons/loaders/GLTFLoader.js';
import { MeshoptDecoder } from 'three/addons/libs/meshopt_decoder.module.js';

const loader = new GLTFLoader();
loader.setMeshoptDecoder(MeshoptDecoder);

const gltf = await loader.loadAsync('/models/chest.opt.glb');
const chest = gltf.scene;
chest.position.set(0, 0, -3); // origin is at the bottom, so y = 0 puts it on the floor
scene.add(chest);

Blender works in metres, and so do most three.js scenes, so the chest should load at about 1 m wide. If it comes in lying on its side, the export probably ran without export_yup. If it's black, your scene needs lights: add a HemisphereLight or an environment map.

6. Iterate

Keep the Blender file open and ask for changes, such as "make the lid open 40 degrees as a separate object named Lid" or "add a gold pile inside". Then run the export and the optimize command again. Ask the agent to keep the export script in your project, so you get the same settings every time.

Limits and safety

  • It runs any Python inside Blender. The README's warning: "The execute_blender_code tool allows running arbitrary Python code in Blender, which can be powerful but potentially dangerous." Python in Blender can read and write any file your user account can. Ask the agent to show scripts before running them, and read them.
  • Save your file before each session. The same warning ends: "ALWAYS save your work before using it." Save versions as you go (File → Save Incremental).
  • Turn on safe mode if you don't need file or network access. Set BLENDER_MCP_SAFE_MODE=1 for the server. Every script is then checked before it runs, and code that reads or writes files directly, runs other programs, uses the network or keeps running after the script ends gets blocked. Modeling, materials, rendering, saving, and import/export still work.
  • Keep it on localhost. The add-on's socket has no authentication or encryption. Anyone who can reach that port can run Python in your Blender. Don't expose it to a network you don't trust.
  • Break big jobs into steps. The README says complex operations may need smaller steps, and timeouts are fixed by simplifying the request. Poly Haven downloads run on Blender's main thread, so Blender freezes until they finish.
  • Check asset licences. Poly Haven is CC0. About 69% of Poly Pizza is CC-BY, which requires credit. The add-on writes the credit line onto imported objects, or you can search with licence="CC0".
  • Telemetry. By default the server sends an anonymous usage record: install ID, session ID, tool name, success, duration, versions, OS and timestamp. Your prompts, code, screenshots and scene data are only collected if you opt in. DISABLE_TELEMETRY=true turns it all off.

Alternatives to Blender MCP

  • Blender Lab's Blender MCP: Blender's own server. It's smaller and built around Python API docs, scene and file summaries, screenshots and running code, including a background mode for .blend files. It needs Blender 5.1+ and is GPL-3.0. Its wiki sets it up by cloning the repo and running it with uv --directory .../blender_mcp/mcp run blender-mcp. Choose it if you want the official project and don't need asset search or AI model generation.
  • Blender Agent Skills: 28 skills for Blender 5.2 LTS+ (modeling, UVs, baking, rigging, animation, optimization, export) for Claude Code, Codex, OpenCode, Gemini CLI, Copilot and Cursor. MIT. Skills teach the agent how to do the work. An MCP server gives it control of Blender. You can use both.
  • blender-game-skills: Claude Code skills that take concept art to rigged, game-ready 3D assets. MIT.
  • Ready-made models. For common props, a free model is often faster than generating one. See Polyfork's low-poly GLBs.

Our game dev agent skills and MCP servers list covers the Godot, Unity, Unreal and three.js equivalents.

FAQ

Is Blender MCP free?

Yes. MCP for Blender is MIT-licensed and the add-on is free. Its Premium page says every feature works on the free plan and only 3D generation is paid. You can use Meshy, Tripo, Hunyuan3D, Hyper3D Rodin or World Labs with your own API keys, or buy Premium (from $10/month on November 21, 2026) to use them without keys. Blender Lab's server is free under GPL-3.0. You still pay for your AI agent's subscription or API usage.

Is Blender MCP safe?

It is as safe as the code you let it run. The agent executes Python inside Blender, and that Python has your user's access to files. Save your work, read scripts before approving them, use safe mode (BLENDER_MCP_SAFE_MODE=1), and keep the add-on's port on localhost. Install it from the official repo or the project's own install script, not from a copy someone re-uploaded.

Is there an official Blender MCP server?

Yes. Blender Lab publishes one at projects.blender.org/lab/blender_mcp, and Claude's connector directory lists it as "Blender", made by Blender Lab. The better-known MCP for Blender (formerly blender-mcp) is a third-party project that isn't affiliated with the Blender Foundation.

Does Blender MCP work with Claude Code?

Yes. Run claude mcp add blender uvx mcp-for-blender, open Blender with the add-on enabled, and check with claude mcp list. Claude Desktop is set up with a JSON block in claude_desktop_config.json instead.

Does Blender MCP work with Codex?

Yes. Run codex mcp add blender -- uvx mcp-for-blender or add [mcp_servers.blender] to ~/.codex/config.toml. The README also offers a Codex plugin that shows a live Blender viewport next to the chat.

Does Blender MCP work with ChatGPT?

The README lists desktop and coding clients (Claude Desktop, Claude Code, Codex, Cursor, VS Code, Windsurf, OpenCode, Antigravity), not the ChatGPT app. Codex is OpenAI's agent for this. The server talks to a Blender running on your own computer, so you need a client that can launch a local MCP server.

Where is the Blender MCP tab?

In Blender's 3D viewport, press N to open the sidebar. The tab is called MCP for Blender. If it's missing, enable Interface: MCP for Blender in Edit → Preferences → Add-ons.

Blender MCP or a Blender skill?

An MCP server lets the agent act on a live Blender: create objects, look at the viewport, export. A skill teaches the agent Blender workflows and conventions, which helps whether it then works through MCP or by running Blender headless from the command line. For game assets, using both works well. The MCP server does the building, and a skill pack such as Blender Agent Skills covers UVs, optimization and export settings.

Sources

Checked November 21, 2026: mcp-for-blender README, MCP for Blender Premium, Blender Lab blender_mcp and its setup wiki, Claude Code MCP docs, Codex MCP docs, Blender export_scene API, Blender glTF 2.0 manual, glTF-Transform CLI, three.js GLTFLoader.