A conversational remote control for Blender
BlenderMCP puts an AI assistant on one side of a bridge and Blender on the other. A Blender add-on listens on TCP port 9876 and turns tool calls into scene operations.
The useful part is not merely asking for a cube. The tool can inspect objects, lights, and cameras, manipulate geometry and materials, execute Blender Python, and bring in assets from Poly Haven and Sketchfab. Hyper3D Rodin and Tencent's Hunyuan3D add generated models. That combination supports quick scene blocking, variations, repetitive cleanup, lighting experiments, and the first pass at a reference image.
It does not remove the need to understand Blender. A prompt may produce awkward topology, wrong scale, poor composition, or destructive code. An artist still has to judge the scene and know how to repair it. The best framing is an unusually fast junior assistant whose every action can affect the open file.
Two installs, then the troubleshooting starts
The happy path is brief. Install uv, add uvx blender-mcp to the MCP client's configuration, install addon.py through Blender's add-on preferences, enable it, and click Connect in the BlenderMCP sidebar. Python 3.10 and Blender 3.0 are the listed minimums. A pipx route exists for users avoiding uv.
Real setup has more seams. Graphical clients often do not inherit the terminal PATH, so the README explains how to locate uvx and place its full path in configuration. Windows may need a cmd /c wrapper. Conda and pyenv users get a Python 3.11 pinning recipe. Configuration changes require fully quitting and restarting the client, and the guide says not to launch uvx manually because the MCP host owns that process.
Only one MCP server instance should control Blender at once, specifically not Cursor and Claude Desktop simultaneously. Blender also needs to be open with the add-on server running. An open issue reports that starting the MCP integration while Blender is closed can delay the handshake and block other servers in one client until timeout. The README itself admits that the first command often fails and recommends trying again. Those are manageable prototype problems, but poor foundations for unattended automation.
Complex requests have their own timeout risk. The official advice is to split them into small sequential prompts. A July 2026 report describes trivial tool calls completing inside Blender but returning about five minutes later, after the MCP client had already canceled them. Other reports concern incomplete JSON replies and missing viewport screenshots. Test the simple scene-info and screenshot paths before building a workflow around long operations.
Powerful tools need a tight trust boundary
The headline security fact is printed in the README: execute_blender_code runs arbitrary Python inside Blender. That code acts with the Blender process's user permissions, not inside a safe scene-only language. It can alter or delete scene content and may access files or network resources available to the process. Saving before use is the minimum. A separate operating-system account, restricted working directory, and reviewed prompts are sensible for untrusted inputs.
The add-on can also download and unpack outside assets. An open August 2026 report points to Hunyuan-generated ZIP extraction without the path-containment check used by sibling Sketchfab and Poly Haven paths. The reporter traced the exact extraction calls and explains how crafted archive names could write outside the temporary directory. Until that path is fixed and released, security-conscious users should avoid Hunyuan imports or inspect and extract results separately.
Telemetry is enabled unless disabled through an environment variable or the add-on preference. The README explains both controls but does not detail the event fields in its security section. Organizations with privacy requirements should disable it during evaluation and inspect network behavior before approval. Optional Sketchfab, Hyper3D, and Hunyuan features require credentials stored in add-on preferences or environment variables, creating more secrets to protect.
Remote host support expands the design beyond one workstation, but it also changes the threat model. The MCP server talks JSON over a plain TCP socket to the add-on. Do not expose that code-capable listener broadly merely because host and port environment variables make it possible. Place any remote route inside an authenticated private network and restrict who can reach it.
Integrations add reach and fragility
Poly Haven can supply HDRIs, textures, and models without leaving the prompt workflow. Sketchfab adds searchable community assets. The two generation services can turn text or images into imported geometry. These are useful shortcuts, but they bring licensing checks, download size, quotas, service accounts, and API drift into a Blender session.
That drift is visible now. A June 2026 issue says the Hunyuan official mode still calls an older Tencent namespace and actions, while the reporter's valid account succeeds against the newer AI3D endpoints. Credentials can therefore appear broken when the integration itself is outdated. Treat every outside connector as optional and verify it independently before blaming the MCP bridge.
Fast-moving project, young release discipline
The repository was pushed on August 8, 2026, and issues were being discussed in August. GitHub's open count of 93 includes both issues and pull requests. More than 25,600 stars show exceptional attention, and the README links a Discord, wiki, tutorials, and contributor group. This is an active project, not an abandoned demo.
Release discipline is less convincing. The GitHub latest-release API returns no release, even though the README directs users to Releases for a changelog and tells upgraders to replace addon.py plus re-add the MCP server. Without a tagged stable point, pinning a reviewed combination of server and add-on takes extra work. Current troubleshooting and security reports reinforce the need to control updates.
BlenderMCP is worth using when the goal is supervised creative acceleration. Start with a copied scene, keep optional integrations off, test inspection and screenshot calls, then grant more capability gradually. For deterministic production automation, Blender's Python API is safer and easier to review. For exploratory prompting with a human watching, this bridge is genuinely useful despite its young edges.