"Blender MCP" now points to two different projects. The official one is a small server from Blender Lab, part of the Blender Foundation, built to inspect and explain scenes through Blender's Python API and its bundled manual. The popular community one, renamed from blender-mcp to MCP for Blender in September 2026, is built to create: it runs Python in your scene, shows the agent what it made, and pulls in assets and AI generated models. Pick the community server to build scenes from prompts in Claude, Codex or Cursor. Pick the official server to audit, document or debug real production files, especially with a local model.

Both connect the same way: your AI client starts an MCP server, and that server talks over a local TCP socket to an addon running inside Blender. If MCP itself is new to you, start with our plain explanation of MCP.

Blender MCP at a glance: official vs community

ItemBlender Lab MCP Server (official)MCP for Blender (community)
MaintainerBlender Lab, Blender FoundationIndependent developer (ahujasid on GitHub), not affiliated with Blender
LicenseGPL 3.0 or laterMIT
Blender needed5.1 or newer3.0 or newer
Installgit clone, uv run blender-mcp, or an .mcpb bundleuvx mcp-for-blender setup
Tools we counted2610
StrengthScene analysis, bundled API docs and manual, batch work on saved filesBuilding scenes, viewport feedback, assets and 3D generators
Guard railsA light block list only, the docs advise a virtual machineOptional safe mode
TelemetryNone in the serverAnonymous usage record by default, opt out with an env var
Default port98769876

The community server: MCP for Blender

This is the project most people mean when they search for Blender MCP. The GitHub repository had about 34,400 stars on November 14, 2026, it sat on the GitHub daily trending list for Python that morning, and it shipped versions 2.2.8 and 2.2.9 on November 12 and 13. On September 16 the maintainer renamed the package from blender-mcp to mcp-for-blender and moved the repo. Old configs keep working because blender-mcp on PyPI is now a wrapper that installs the new package.

The fast path is one command. If you already have uv, run uvx mcp-for-blender setup. It looks for Claude Desktop, Claude Code, Codex, Cursor, VS Code, Windsurf, OpenCode and Antigravity, adds a blender entry to the ones you pick, keeps a backup of each config file, and installs and enables the addon. Manual setup is just as short: claude mcp add blender uvx mcp-for-blender in Claude Code, codex mcp add blender -- uvx mcp-for-blender in Codex, or a uvx entry in your Cursor or OpenCode config, then uvx mcp-for-blender install-addon. The addon starts its socket server when Blender opens. In our test it was listening on port 9876 within seconds.

With a client that lacks the inline viewport app, the server listed 10 tools whose definitions came to about 20,200 characters of JSON. The ones that matter:

  • execute_blender_code runs any Python in your live scene. Ours added a cylinder and returned its print output.
  • look returns an image of the viewport, the camera view, a multi angle sheet or an animation strip, so the agent can check its own work. Ours came back as a 243 KB image.
  • get_scene_info gives a short text summary, one line per object with location and size.
  • search_assets and import_asset pull from Poly Haven, Sketchfab and Poly Pizza. Each library is off until you tick it in the Blender sidebar, and the tool tells the agent exactly that.
  • generate_3d and generate_gaussian_world call Tripo, Hunyuan3D, Hyper3D Rodin, Meshy or World Labs with your own API keys, or through the paid Premium tier without keys.

The Codex plugin adds a live viewport inside the chat and object mentions, but it is not in the Codex plugin directory yet, so you add it from the repo.

The official server: Blender Lab MCP

Blender Lab published its MCP Server page with a different goal: "a natural language interface with Blender's Python API", aimed at understanding and fixing complex files. Its own examples are telling. It finds polygon count outliers in the Classroom benchmark scene, renames data blocks, answers which objects use a material, and documents a geometry nodes setup with frames and a text block. The page also admits the first answer missed a Solidify modifier, which is a fair reminder to check numbers an agent gives you.

Setup takes more steps. You need Blender 5.1 or newer, the MCP extension (drag it into Blender twice: once to add the Blender Lab repository, once to install it), a client, and the server. The setup wiki has you clone the repository and point your client at uv --directory <clone>/mcp run blender-mcp. Clients that read .mcpb bundles can install a packaged release instead. The examples use llama.cpp, so a local model is a first class path here.

The server listed 26 tools, about 20,600 characters of definitions. Beyond execute_blender_code you get:

  • Summaries of the open file: data block counts, missing external files, linked libraries, path and save state, and a guess at what the file is for.
  • Object summaries and per object detail (we got location, dimensions and modifiers back for Suzanne).
  • Navigation tools that jump the interface to an editor tab or an object.
  • Window and area screenshots, a thumbnail render and a viewport render.
  • search_api_docs, search_manual_docs and get_python_api_docs, which search an offline copy of the Blender 5.2 API reference and manual shipped with the server. Our query for bevel width returned the right BevelModifier entries.
  • Six _for_cli twins that open a saved .blend in background Blender, no window needed. We pointed BLENDER_PATH at Blender 5.3 and got data block counts and a custom Python result from a file nobody had open.

A few details from testing and the source. The viewport render took over a minute on our machine with no GPU, and it saved into Blender's temporary folder under the file name we passed, not the full path we asked for. And there is no real safe mode. The addon blocks a short list of calls such as sys.exit, and its source calls this a weak sandbox that a motivated model can work around. The Blender page says it runs model written code "without any guards in place" and recommends a virtual machine.

Running both: the port 9876 clash

Both addons listen on localhost port 9876 by default, and both servers connect there unless told otherwise. We started Blender with the community addon, then a second Blender with the official extension. The second one could not take the port and printed nothing about it. When we then called an official tool, the official server reached the community addon instead, got no answer it understood, and hung for more than two minutes until we stopped it.

The fix is to give one pair a different port. Set the port in the addon's preferences, then match it on the server side. The two projects use different variable names: BLENDER_MCP_PORT and BLENDER_MCP_HOST for the official server, BLENDER_PORT and BLENDER_HOST (or a --port flag) for the community one. Also watch the command names: the official server's command is blender-mcp, which is also the community project's old PyPI name. uvx blender-mcp installs the community wrapper, never the official server. Our guide to running MCP servers locally covers checking which process answers on a port.

Safety: what each one lets the agent do

Both servers give the model a Python interpreter inside Blender, and Blender's Python can read files, open network connections and start programs. Save your work first, as the community README warns in bold.

The community server's safe mode is worth turning on for anything beyond toy scenes. Add BLENDER_MCP_SAFE_MODE=1 to the server's environment. In our test it rejected a script that imported os and sent the reason back so the agent could rewrite it, while normal modeling code still ran. Rendering, saving and import or export stay allowed. Its remote access option has no authentication or encryption, so keep it on localhost or use an SSH tunnel.

On telemetry, the community server sends a minimal anonymous usage record by default (an install ID, tool names, success, timing, versions and operating system). Prompts, code and screenshots are only collected if you opt in. Set DISABLE_TELEMETRY=true to stop the usage record too. The official server sends nothing of its own. This is the same official versus community split we covered for the Home Assistant MCP server: the vendor version is narrower and more conservative, and the community version does more.

Which Blender MCP should you use?

  • Making things from prompts (a low poly scene, a product shot, set dressing with Poly Haven assets): MCP for Blender. Its look tool and asset libraries are what make prompt based modeling workable.
  • Working on real production files (finding heavy objects, missing textures, broken library links, naming cleanup, explaining someone else's node tree): the Blender Lab server. The docs search and background tools fit that work, and it is maintained by the people who make Blender.
  • Local or offline models: the Blender Lab server, which documents llama.cpp, and whose bundled docs keep a smaller model from guessing at API names.
  • Old Blender versions: MCP for Blender, which supports 3.0 and up.

For an agent that also needs a browser while it works, see our Chrome DevTools MCP guide.

FAQ

Is Blender MCP free?

Yes. Both servers and their addons are free and open source, the official one under GPL 3.0 or later and the community one under MIT. You pay for your AI client or model. The community project also sells an optional Premium tier for generating 3D models without your own Meshy, Tripo, Hunyuan3D, Rodin or World Labs keys. See its site for current prices.

Is the Blender MCP addon made by Blender?

Only one of them. The Blender Lab MCP Server comes from the Blender Foundation. MCP for Blender, formerly blender-mcp, is a third party project that says it is not made by Blender. The rename was partly to make that clearer.

How do I use Blender MCP with Claude Code or Codex?

For the community server, run claude mcp add blender uvx mcp-for-blender or codex mcp add blender -- uvx mcp-for-blender, install the addon with uvx mcp-for-blender install-addon, and open Blender. For the official server, add a stdio entry that runs uv --directory <clone>/mcp run blender-mcp after cloning the Blender Lab repository and installing its extension.

Why does my Blender MCP server say it cannot connect?

The addon's socket server is not running or is on another port. Check the MCP for Blender tab in the sidebar (press N) or the official extension's preferences, and click start if needed. If both addons are installed, only one can hold port 9876. On desktop apps a spawn uvx ENOENT error means the client cannot find uvx; use its full path.

Can I run Blender MCP with a local model?

Yes. Any client that speaks MCP works, and Blender Lab documents llama.cpp. Small models do better with the official server because it can look up the Blender API in its bundled docs instead of guessing.