Blender MCP: Driving 3D Through Claude (2026)

Q: How safe is this for production projects?

Handle with care. Because executeblendercode has no sandbox, don't run the setup on untrusted files, and always save beforehand. For production, isolate the environment (separate user account or container) and review the code.

In short

Blender MCP is a bridge between the Blender 3D editor and Claude: you type "add a table and three chairs, set up soft lighting," and Claude executes the commands in your open scene. Technically it's two pieces — an add-on inside Blender that runs a socket server, and a Python MCP server that translates the model's requests into Blender Python API calls. As of July 2026 there are two independent projects: the community server ahujasid/blender-mcp (open source, works with any LLM) and the official Blender Lab server, which Anthropic released as a Blender connector in spring 2026. Both are free. Below: what the setup can actually do, how to get it running in 10 minutes, and where the security landmines are.

What Blender MCP is and how it's built

MCP (Model Context Protocol) is an open standard for connecting a language model to external tools and data. Blender MCP applies that standard to 3D: it gives Claude "hands" inside Blender.

The architecture is simple, and worth understanding because debugging depends on it:

The Blender add-on (addon.py) — runs a TCP socket server inside Blender itself (port 9876 by default) and listens for JSON commands.
The Python MCP server — launched via uvx blender-mcp, implements the MCP protocol for Claude and sends commands into the add-on's socket.

The exchange uses plain JSON objects: a command with a type and parameters → a response with status and result. If you've built something similar before, take a look at the walkthrough on building an MCP server in Python — same logic, just with the far end of the pipe plugged into bpy.

Community server or official Blender Lab: which to pick?

As of July 2026 the query "blender mcp" has two answers, and it matters not to confuse them.

ahujasid/blender-mcp — the project that started the hype. Open source, "Blender with any LLM," rich integrations (Poly Haven, Hyper3D Rodin, Sketchfab), installs in a couple of clicks. This is the one most guides and videos are about.

The official Blender Lab server is a Blender connector Anthropic released in April 2026 (the repository lives in Blender Lab infrastructure; release v1.0.0 came on 27 April 2026, and the add-on requires Blender 5.1). Around the same time Anthropic sent Blender roughly €240,000: it was first announced as corporate patron membership in the Blender Development Fund, but after community discontent over generative AI the Blender Foundation re-designated the money in May 2026 as a one-off unconditional donation with no formal membership (per CG Channel). The official server's focus isn't "generate me a model" but a careful natural-language interface to the Python API: analysing and debugging scenes, editing objects in bulk, answering documentation questions.

The practical takeaway: for experiments and asset generation, take the community server; for clean work on existing scenes and scripts, look at the official one. More ready-made servers are in the MCP server roundup and in the /arsenal/mcp/ section.

What Blender MCP can do: capability table

Below are the capabilities of the community server ahujasid/blender-mcp (as of July 2026). The "Key/service" column shows whether external access is needed.

Capability	What it does	Key/service
Scene inspection	Returns the object list, their properties, the hierarchy	No
Creating and editing objects	Creates, deletes, moves and modifies primitives and meshes	No
Materials	Creates and assigns materials and colours	No
`execute_blender_code`	Runs arbitrary Python directly in the Blender session	No
Viewport screenshot	Returns a viewport image so the model can "see" the scene	No
Poly Haven	Pulls models, textures and HDRIs from the library	Free, no key
Hyper3D Rodin	Generates a 3D model from text or an image and places it in the scene	Free-trial key, daily limit
Sketchfab	Searches and downloads models	Sketchfab account

The key and most powerful tool is execute_blender_code. execute_blender_code is the tool that grants the model the right to run any Python inside Blender. It's how you do everything the "ready-made" commands don't cover. It's also the main source of risk (see below).

On the services: as of July 2026 Poly Haven serves free HDRIs, textures and models with no API key; Hyper3D Rodin offers a free-trial key with a daily generation limit, and full keys come from hyper3d.ai and fal.ai.

Connecting Blender MCP to Claude and Cursor, step by step

You'll need Blender 3.0+ and Python 3.10+, plus the uv package manager. The sequence (for the community server):

Install the add-on. Download addon.py from the repository, then in Blender: Edit → Preferences → Add-ons → Install → pick the file. Tick "Interface: Blender MCP."
Start the socket server in Blender. In the 3D viewport press N, open the "BlenderMCP" tab in the side panel and hit Start. You'll see a message that the server is up on port 9876.
Register the MCP server with your client. For Claude Desktop, add this to the config:

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

For Cursor — Settings → MCP with the same command (on Windows, wrap it: cmd /c uvx blender-mcp). The same server attaches to Claude Code too — exactly the mechanics described in the guide on connecting MCP to Claude Code.

Test the connection. A Blender tools icon will appear in the client. Give it a simple prompt: "Describe the current scene." If Claude returns an object list, the setup works.

One important nuance: keep only one client active. If the MCP server is running in both Cursor and Claude Desktop at once, they'll fight over the socket — pick one.

What are the limitations and risks of Blender MCP?

Honestly: it's a powerful tool that is still raw on the security side. What to keep in mind (based on the project's issue tracker and Blender DevTalk discussions through 2026):

Arbitrary code with no sandbox. execute_blender_code runs Python via exec() with no isolation. That code has your permissions: file system, network, launching processes. ALWAYS save your work before starting the setup.
Prompt injection. A malicious instruction can be embedded in a .blend file, a scene description or any text the model reads. Don't feed the setup projects or prompts from untrusted sources.
Some tools touch files. A few operations (viewport screenshots, image-based generation) read arbitrary paths on disk — keep the project in a separate, clean folder.
Telemetry is on by default in the community server; turn it off with the DISABLE_TELEMETRY=true variable.
Non-determinism. The model sometimes hallucinates bpy operator names and the command fails; on complex scenes a targeted script often beats a free-form prompt.

Practical mitigation: run it under a separate user account or in a container, don't give it access to your working directories, and review generated code before running it.

How to fold Blender MCP into vibe coding

Vibe coding is the approach where you describe the task in words and the agent writes and runs the code for you. Blender MCP is exactly the same move, except here the "code" is a 3D scene. If you already work in Claude Code or Cursor, adding the Blender server extends your usual loop into graphics: same chat, same tools, plus hands in 3D. On the approach itself, see the explainer on what vibe coding is.

To keep the setup from living in a vacuum, it helps to have a configured agent environment: rules, prompts, MCP wiring. The Quest engine provides exactly that framework for Claude Code — and Blender MCP plugs into it as one more server from the MCP arsenal.

FAQ

Do I need a paid Claude plan to use Blender MCP?

Blender MCP itself is free and open source — you only pay for model access. The setup works with any MCP-compatible client (Claude Desktop, Claude Code, Cursor). Claude access pricing changes, so check current terms with the provider as of July 2026.

Can Claude generate a 3D model from scratch?

Partly. The Hyper3D Rodin integration gives you a rough blank from text or an image, and Poly Haven and Sketchfab pull ready-made assets. But precise retopology, UV unwrapping and production geometry are still finished by a human — the model is good as a draft and as an operator for routine work.

Does Blender MCP work with Cursor, not just Claude?

Yes. The README has configs for both Cursor and Claude Desktop; the MCP protocol is the same. One rule: keep only one client active at a time, or they'll fight over socket port 9876.

How does the official Blender Lab server differ from ahujasid/blender-mcp?

The official one (v1.0.0, 27 April 2026, for Blender 5.1) focuses on scene analysis and debugging and on a careful interface to the Python API. The community server is broader on "creative" integrations (generation, assets). For generation, go community; for engineering work on scenes, go official.

How safe is this for production projects?

Handle with care. Because execute_blender_code has no sandbox, don't run the setup on untrusted files, and always save beforehand. For production, isolate the environment (separate user account or container) and review the code.