Blender MCP Server

The Blender MCP Server makes the 3D software Blender accessible to AI assistants. Instead of clicking through the outliner, data-blocks, and the Python reference, you ask questions in plain language: Which objects use this material? Where does my scene carry needless geometry? Are any linked files missing? The server translates such requests into calls to the Blender Python API and returns structured answers. It is maintained by Blender Lab, the innovation space of the Blender project, which makes it the official solution rather than one of the many community bridges.

What is the Blender MCP Server?

According to the provider, the project is deliberately small and does no more than necessary. It consists of two parts that talk over a TCP socket: an add-on that runs inside Blender and executes requests, and an MCP server that the MCP client launches as a separate process and talks to over stdio. The server also bundles excerpts from the Blender Python API reference and the user manual, so the model can work with documentation at hand.

Features and tools

The tool list in the repository is compact and easy to follow. Analysis covers summaries of a blend file's data-blocks, a check for missing external files, a tree of linked libraries, path and save information, and a guess of what the file is mainly used for. On top of that come the scene hierarchy, detail summaries of single objects, screenshots of a window or area, and a JSON description of the window layout. The model can focus the 3D viewport on an object, switch workspaces, and write small preview or viewport renders. Full-text search over the API reference and the manual lets it find suitable functions itself. Many analysis tools also exist as a variant for background operation without an open interface.

The most powerful, but also most delicate tool is execute_blender_code: it runs model-generated Python code inside the running Blender instance.

Setup

According to the provider you need Blender 5.1 or newer. You download the add-on as a zip file of the current release and drag and drop it into Blender, twice according to the instructions: once for the repository, once for the add-on itself. You install the MCP server either as an .mcpb bundle for newer clients or from source. llama.cpp is explicitly documented as a client; because the server follows the MCP standard, other clients that can launch local stdio servers are expected to work too. In the add-on you can set host, port, and an optional auto-start.

Security: code runs without protection

The provider warns clearly: the MCP server executes LLM-generated code in Blender without any guards in place to protect your data. Isolated virtual machines or systems without sensitive information are recommended. Because Blender Python has full file system access, take this seriously, work on copies of your projects, and do not hand out blanket approvals in your client.

Licence and cost

According to the LICENSE file in the repository, the source code is released under the GNU General Public License version 3. The server itself is free and open source; costs can only arise from the language model or client you use.

Who is the Blender MCP Server for?

It is aimed at Blender users and technical artists who want to understand, tidy up, and document large or inherited scenes, for example unify naming, describe geometry-nodes setups, or find performance bottlenecks. Add-on developers benefit too, because the model looks up the matching API documentation directly. If you just want to model a few objects, you will get there faster without a server; if you want a project analysed or routine tasks scripted, it saves a lot of searching.

Blender 5.1 or newer with the Blender MCP add-on installed and enabled, plus an MCP client that can launch local stdio servers (llama.cpp is documented).

The official Blender MCP Server from Blender Lab connects an MCP client to a running Blender instance. This guide shows the path to a first working connection and what to watch out for along the way.

Step 1: Check the prerequisites

You need Blender 5.1 or newer and an MCP client that can launch local stdio servers. llama.cpp is explicitly documented on blender.org/lab/mcp-server.

Step 2: Install the add-on

Download the zip file of the current release from blender.org/lab/mcp-server and drag and drop it into Blender. According to the instructions this is needed twice: once for the repository, once for the add-on. Then enable the add-on in the preferences.

Step 3: Connect the MCP server

Install the server as an .mcpb bundle or from source at projects.blender.org/lab/blender_mcp and register it in your client as a stdio server. The entry point is called blender-mcp. You set host and port of the connection to Blender in the add-on preferences, where an optional auto-start also lives.

Step 4: Test a first request

Start with a read-only question, such as a summary of the blend file or a search for missing files. That checks the connection without running code that changes your scene.

Contain code execution

The execute_blender_code tool runs model-generated Python code without guards. Work on copies of your projects, approve calls in your client one by one, and use an isolated virtual machine for sensitive data.

No sign-in. The server connects to a locally running Blender instance over a configurable TCP host and port.

Read access to scene, data-blocks, and file paths, plus execution of arbitrary Python code in Blender with the permissions of the user who runs Blender.

The server runs locally. Scene data, screenshots, and documentation excerpts enter the client context and therefore reach the respective model provider; with a local model (e.g. via llama.cpp) they stay on your own machine.

According to the provider, the server runs model-generated code in Blender without guards; Blender Python can access the file system. Isolated virtual machines or systems without sensitive data, working on project copies, and no blanket approvals are recommended.