Blog · Coding agents

Blender MCP: install it in Claude Code and Codex

Blender MCP is an MCP server plus a Blender add-on that lets an AI agent such as Claude Code or Codex control a running copy of Blender, mostly by writing and running Blender Python in your open file. The name usually means mcp-for-blender, a community project by Siddharth Ahuja (formerly blender-mcp); Blender Lab, part of the Blender project, also publishes a smaller official server. This guide explains how the two differ, gives the install commands for Claude Code and Codex, walks through a short example where an agent builds a low-poly prop and exports a GLB, and covers what to lock down before you let it run. Updated November 21, 2026.

What Blender MCP is

MCP (Model Context Protocol) gives an agent tools that act on another program. Both Blender MCP projects have the same two parts: an add-on that runs a small server inside Blender, and an MCP server that your agent starts and that forwards its tool calls to the add-on as JSON over a TCP socket.

The project most people mean is mcp-for-blender. Its README opens with a disclaimer that it is a third-party integration and not made by Blender. It is MIT-licensed, had 35,153 GitHub stars on November 21, 2026, and its PyPI package was renamed from blender-mcp to mcp-for-blender (version 2.3.1 on November 18); uvx blender-mcp still starts the server, so older configs keep working.

Blender Lab's blender_mcp is the official one. Its readme calls it a lightweight server, deliberately small, aimed at a natural-language interface to Blender's Python API, better access to the documentation, and exploring complex scenes. It is listed in Claude's connector directory as Blender, made by Blender Lab, Anthropic verified.

mcp-for-blender vs Blender Lab's blender_mcp

Figures from the GitHub and projects.blender.org repositories on November 21, 2026.

mcp-for-blender (community)blender_mcp (Blender Lab)
Made bySiddharth Ahuja; not affiliated with the Blender FoundationBlender Lab, part of the Blender project
LicenceMITGPL-3.0
Blender version3.0 or newer5.1 or newer
Stars35,153 on GitHub39 on projects.blender.org
ToolsNine, including execute_blender_code, look (viewport and camera views), get_scene_info, asset search and import (Poly Haven, Sketchfab, Poly Pizza) and generate_3d (Tripo, Hunyuan3D, Hyper3D Rodin, Meshy)execute_blender_code, execute_blender_code_for_cli (background Blender), blend-file and object summaries, get_python_api_docs, screenshots; bundles Python API and manual excerpts
InstallOne-command installer or uvx mcp-for-blender; built-in setup for Claude Code, Codex, Cursor and othersgit clone the repo and run it with uv; add the add-on in Blender
Best forBuilding and dressing scenes, pulling in assets, generating modelsInspecting files and asking about Blender's API, on current Blender

Quick install: all your agents at once

The mcp-for-blender README's recommended route is one command. It installs uv if needed, finds the MCP clients on your machine (Claude Desktop, Claude Code, Codex, Cursor, VS Code, Windsurf, OpenCode, Antigravity), lets you pick which to set up, then installs and enables the Blender add-on. It only adds a blender entry to each config and keeps a .bak of every file it changes. Afterwards, fully quit and reopen your agent.

The commands below are copied from the README; we didn't run them. Piping a script into your shell runs it with your permissions, so if you'd rather see each step, use the manual setup in the next two sections instead.

  • 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
  • With uv already installed: uvx mcp-for-blender setup
  • You need Blender 3.0 or newer, Python 3.10 or newer, and uv.

Install Blender MCP in Claude Code

Manual setup for mcp-for-blender. Commands are copied from the project README and the Claude Code MCP docs; we didn't run them.

  1. Install uv

    uv provides the uvx command that starts the server. The README warns against pip install uv; use the official installer.

    # macOS
    brew install uv
    
    # Linux
    curl -LsSf https://astral.sh/uv/install.sh | sh
    
    # Windows
    winget install --id=astral-sh.uv -e
  2. Register the server

    This saves the server for the current project. Add --scope user to use it everywhere; options like that go before a -- that separates them from the server command, which is the form the Claude Code docs use. Inside a session, /mcp shows the same status.

    # the README's command
    claude mcp add blender uvx mcp-for-blender
    
    # check it: should show the server as connected
    claude mcp list
    claude mcp get blender
  3. Optional: turn telemetry off

    The README says to set DISABLE_TELEMETRY=true in the server's environment to turn telemetry off completely. Claude Code's claude mcp add takes environment variables with --env; see the Claude Code MCP docs for the exact form.

  4. Install and enable the Blender add-on

    Then in Blender open Edit → Preferences → Add-ons and enable "Interface: MCP for Blender". Press N in the 3D viewport and open the MCP for Blender tab; the server starts with Blender, and if it isn't running, click Start MCP Server.

    # copies the add-on into Blender's add-ons folder as blender_mcp.py (keeps a .bak)
    uvx mcp-for-blender install-addon
    
    # lists the add-on folders it found
    uvx mcp-for-blender addon-paths

Install Blender MCP in Codex

The Codex CLI, desktop app and IDE extension share ~/.codex/config.toml, so one setup covers all three. Commands are copied from the README and OpenAI's Codex MCP docs; we didn't run them. Install uv and the Blender add-on as in the Claude Code steps.

  1. Add the server with the CLI

    Codex shows the tools the next time it starts; /mcp in the Codex terminal UI lists active servers. In the desktop app the same thing is Settings → MCP servers → Add server, name blender, STDIO, command uvx mcp-for-blender.

    # the README's command
    codex mcp add blender -- uvx mcp-for-blender
    
    # check it: blender should show as enabled
    codex mcp list
  2. Or edit config.toml

    The [mcp_servers.blender] table is from the README. The env line follows the Codex docs, which list env, startup_timeout_sec (default 10) and tool_timeout_sec (default 60) as table keys; raise tool_timeout_sec if long Blender operations time out.

    # or by hand in ~/.codex/config.toml (shared by the Codex CLI, desktop app and IDE extension)
    [mcp_servers.blender]
    command = "uvx"
    args = ["mcp-for-blender"]
  3. Optional: the Codex plugin

    The README recommends this for Codex: it shows a live Blender viewport inside the chat, thumbnail pickers for assets, 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 and install "MCP for Blender" from the plugins directory

Installing Blender Lab's official server instead

Blender Lab's setup wiki has you install its add-on in Blender 5.1 or newer, clone the server into your home folder, and point your MCP client at it with uv. This is the wiki's clone command and config as recorded; we didn't run it. If you use Claude, the server is also listed in Claude's connector directory, which notes that the add-on is required.

git clone https://projects.blender.org/lab/blender_mcp.git, then {"mcpServers": {"blender": {"command": "uv", "args": ["--directory", "$HOME/blender_mcp/mcp", "run", "blender-mcp"]}}}

Don't register both projects under the same blender name. The mcp-for-blender README also says to run only one instance of its server at a time, for example in Cursor or Claude Desktop but not both.

Example: an agent builds a low-poly prop and exports a GLB

With Blender open and the add-on running, the agent does the modelling through execute_blender_code and checks its work with look. We wrote this from the docs and didn't run it end to end, so results will vary. For compressing the GLB with glTF-Transform and loading it in three.js, see the full Blender MCP guide on vibegamer.io.

  1. Save your .blend file

    The agent edits your open file directly. The README's warning is to always save your work before using execute_blender_code.

  2. Prompt the agent

    Give sizes, a triangle budget and the export path. Plain base colours export cleanly, since Blender's glTF exporter reads base colour from the Principled BSDF node. The README also notes complex operations may need to be broken into smaller steps.

    Use the Blender MCP tools. Build a low-poly wooden crate for a browser game:
    - 1 m cube, sitting on the ground at z = 0, origin at the bottom centre.
    - Bevelled planks and corner braces, under 800 triangles.
    - Principled BSDF base colours only (wood, dark metal), no image textures.
    - Apply modifiers, join everything into one object named "Crate".
    - Check it with the look tool from a few angles and fix anything floating.
    - Select only "Crate" and export it as GLB to ./public/models/crate.glb.
    Show me each script before you run it.
  3. Read the export script before it runs

    The export usually ends in a call like this, using parameters from Blender's export_scene API. GLB packs mesh, materials and textures into one file. Check that the path is inside your project.

    import bpy
    
    bpy.ops.export_scene.gltf(
        filepath="/path/to/your-game/public/models/crate.glb",
        export_format='GLB',  # one binary file with mesh, materials and images
        use_selection=True,   # only the selected "Crate"
        export_apply=True,    # apply modifiers on export
        export_yup=True,      # +Y up
    )

Is Blender MCP safe? What to lock down

Treat it like giving the agent a Python shell inside Blender. These points are from the mcp-for-blender README.

  • Arbitrary Python: execute_blender_code runs any Python the agent writes. The README calls this potentially dangerous and says to always save your work first. Read scripts before approving them.
  • Safe mode: set BLENDER_MCP_SAFE_MODE=1 to validate scripts before they run. It blocks reading and writing files directly, running other programs, network access and persistent code; modelling, materials, rendering, saving and import/export still work.
  • No socket authentication: the add-on's server (port 9876 by default) has no authentication or encryption, so anyone who can reach that port can run Python in Blender. Keep BLENDER_HOST on localhost unless you're on a trusted network.
  • Telemetry: by default it sends an anonymous install ID, session ID, tool name, success, duration, versions, OS and timestamp. Prompts, generated code, screenshots and scene data are only sent if you opt in, and opted-in data may be used to train AI models. Set DISABLE_TELEMETRY=true to turn it off entirely.
  • Third-party keys: 3D generation and some asset sources use API keys passed as environment variables (for example BLENDERMCP_MESHY_API_KEY). Poly Pizza models are about 69% CC-BY, so credit is required unless you filter to CC0.

Blender MCP or a Blender skill?

An MCP server gives the agent a live connection to a running Blender. A skill gives it written instructions and scripts for a workflow, such as EL4CTEO/blender-skills (28 Blender 5.2 LTS+ workflow skills for Claude Code, Codex and others). They work well together: the skill tells the agent how to build a game-ready asset, the MCP server lets it do that in your open scene.

We list Blender and game-making skills and MCP servers in game development skills, and vibegamer.io compares them in game dev agent skills.

Crewmark: a 3D artist on the same board as your coders

Blender MCP is built around one session at a time: you prompt, the agent models, you review. If you want an agent that keeps making assets over days while others write the game code, Crewmark is another option.

In Crewmark each agent is an AI employee: a long-running session with a role, running on your own machine. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx). Work moves through a shared issue board, where you and the employees assign issues, comment, @mention each other and change statuses. That's how a coding employee asks a Blender employee for a new prop: it @mentions it on the issue. Employees wake on a schedule (heartbeats) and whenever they're mentioned, so the work starts without you passing it along.

It doesn't replace Blender MCP; an employee with the server installed still uses it. It adds the roles, the board and the wake-ups around it.

FAQ

Is Blender MCP free?

Yes. mcp-for-blender is open source under MIT, and its site says every feature works on the free plan except 3D generation, which needs a Premium plan (from $10 a month) or your own provider API keys. Blender Lab's server is free under GPL-3.0.

Is Blender MCP safe?

It runs any Python the agent writes inside Blender, and its socket server has no authentication. Save your work first, read scripts before approving them, keep it on localhost, and set BLENDER_MCP_SAFE_MODE=1 to block file access, other programs and network access. Set DISABLE_TELEMETRY=true to turn off usage telemetry.

Is there an official Blender MCP server?

Yes. Blender Lab publishes blender_mcp on projects.blender.org (GPL-3.0, Blender 5.1+), and it's listed in Claude's connector directory as made by Blender Lab. The popular mcp-for-blender is a third-party project not affiliated with the Blender Foundation.

How do I add Blender MCP to Claude Code?

Install uv, run claude mcp add blender uvx mcp-for-blender, check it with claude mcp list, then install and enable the Blender add-on with uvx mcp-for-blender install-addon. Or run the README's one-command installer, which sets up Claude Code for you.

How do I add Blender MCP to Codex?

Run codex mcp add blender -- uvx mcp-for-blender, or add a [mcp_servers.blender] table with command = "uvx" and args = ["mcp-for-blender"] to ~/.codex/config.toml. The CLI, desktop app and IDE extension share that file. The README also offers a Codex plugin with a live viewport in the chat.

Should I use Blender MCP or a Blender skill?

Use MCP when the agent needs to act on a running Blender scene; use a skill when you want to give it a tested workflow to follow. Many setups use both: the skill describes the workflow and the MCP server carries it out.

Commands and features checked against the mcp-for-blender README, the Blender Lab blender_mcp repository and wiki, the Claude Code and Codex MCP docs, and Blender's glTF export docs on November 21, 2026. We didn't run the install commands.

Give your game a 3D artist that works with your coders.

Crewmark gives each agent a role, a shared issue board and scheduled wake-ups, so they hand work to each other without you copying it between terminals.

Crewmark is open source: star us on GitHub