Blender MCP Server on Windows: Claude AI Setup, Animation Tips and Security Risks

The story is encouraging, but it needs some context. The integration is real, official and fairly simple to set up on a Windows PC. It is also a direct route for AI-generated Python code to run inside Blender, and Blender's own documentation is blunt about what that means for your files.

What Ghosh Says He Built​

Ghosh's account is a personal experiment, not a benchmark. Here is what he reports:

• The result: a 45-second animated clip made in a session of about four hours. He says he was actively involved for less than one of those hours, mostly reviewing Claude's work and restating what he wanted.
• The setup: Claude Opus 5.5 in low-effort mode on the $20 Pro plan. He says the project used about 26% of his session allowance.
• The process: the published clip was the third attempt. Earlier versions had trouble with camera movement, character models and scenery depth.
• The flaws: he lists them himself. The robot floats above the ground, its head passes into a flower when it touches it, and he describes the river colors in unflattering terms.

His most useful finding is about how to structure the request. He first asked Claude to open on a small area and zoom out to show the whole world. Each time the camera pulled back, the scene got sparser. Claude wasn't building the world from the ground up and seemed to struggle with creating everything at once. He then reversed the order: build the whole world in as much detail as possible first, then choose a spot and zoom out from there. He compares it to drawing the map and writing the lore before working on the characters.

That advice is sensible, but it's one person's observation. Nobody has shown that it works reliably across projects. Treat the time, the usage figure and the quality as one user's experience, not something you can expect to repeat.

Section summary: a beginner got a watchable but flawed short animation after three rounds of revision. The most transferable tip is to build the whole scene before you frame the shot.

What the Official Blender MCP Server Actually Is​

Blender's documentation fills in what the personal account leaves out. The Blender Lab page describes a lightweight MCP (Model Context Protocol) server for Blender. It offers a natural language interface with Blender's Python API, improving access to documentation, and allowing users to explore and understand complex setups.

The project's readme is more specific. It says the server allows LLM assistants to interact with a running Blender instance – inspecting scenes, executing Python code, rendering, and navigating the interface. It supports running arbritary Python code within Blender. The readme also says the server bundles the complete Python API reference and user manual, so the model can look up current documentation for both.

A few more facts that matter:

• It isn't built into Blender. Blender states that Blender does not have any built-in functionality for connecting to LLMs. In order for Blender to connect with LLMs, three external tools must be manually downloaded, installed, and run. Those are the add-on, an LLM client and the MCP server. You need Blender 5.1 or newer.
• It's a Lab project, not part of core Blender. A tool directory, mcp.film, calls it official but experimental — a Blender Lab project, explicitly outside the current Blender roadmap. Blender's Lab page lists the MCP Server as released, but also notes that Lab activities are not initially part of the roadmap and have no release timeline or target.
• Blender's own examples are not about making animations. The official page shows scene analysis, such as finding objects with high polygon counts relative to their size on screen, fixing typos in data-block names, querying which objects use a given material, and documenting Geometry Nodes setups. Building a whole animated world, as Ghosh did, goes well beyond anything Blender demonstrates. Blender also says success depends on the model you use.
• Watch out for the name clash. A separate community project also called "blender-mcp" predates the official one. One developer's guidance for supporting the Blender Lab version warns that both upstream projects use the name blender-mcp; install the Blender Lab source in a separate environment and use the client name blender-lab. Make sure you download from Blender, not from a search result with a similar name.

Section summary: the server is an official, experimental Blender Lab add-on plus server. It lets an assistant read your scene and run Python in it. It doesn't come with Blender, and Blender doesn't pitch it as an automated animation studio.

The Security Warning You Shouldn't Skip​

This is the most important thing in this piece, and it isn't in Ghosh's account. Blender's MCP page says: The MCP server will execute LLM generated code in Blender without any guards in place to protect your data from removal or being sent to a remote location. Its recommendation is clear: To keep your data safe it is recommended to use a virtual machine, or a system without access to sensitive information.

Blender's 5.1 manual explains why script execution in general is risky. Python places no limits on what a script can do, so you should only run scripts from sources you trust. With MCP, the "source" is a language model writing code as it goes, based on your prompts and whatever is in the scene.

What this means on a Windows PC:

• A separate virtual desktop is not isolation. Ghosh suggests leaving Claude and Blender to work on their own virtual desktop, which is handy for your workflow. But it's the same user account and the same files. Only a VM or a separate machine gives you the separation Blender recommends.
• A VM is the easiest safe option. Hyper-V on Windows 11 Pro, or any hypervisor you already use, gives you a disposable environment. The catch is graphics performance, since Blender relies heavily on the GPU. That's a trade-off you'll have to weigh yourself.
• Keep valuable work out of reach. Test on copies of scenes and back up anything you care about. Don't run it on a machine that has client files, credentials or company data.
• Scrutiny from other developers backs this up. A security audit of a different Blender MCP plugin, a third-party project, cited the MCP specification's rule that servers must validate the Origin header on incoming connections and should bind only to localhost when running locally. Local servers that accept commands from AI agents are getting security review, and they should.

For IT admins, Anthropic's help documentation says Team and Enterprise owners can control which desktop extensions are available, including an allowlist. Machine-level enterprise policy overrides those in-app settings. If your users are sideloading .mcpb files, those are the controls to look at.

Section summary: Blender says outright that the server runs AI-written code with no safeguards. Use a VM or a machine with nothing sensitive on it.

How to Set It Up on Windows​

Ghosh says the setup takes about ten minutes, and his steps match Blender's official sequence. Here is the procedure, with the official details added.

What you need:

• Blender 5.1 or newer
• Claude Desktop, or another MCP client that supports .mcpb bundles
• Ideally, a VM or a machine without sensitive data, as described above

In Blender:

1. Download the MCP add-on from Blender's official MCP Server page. If you install it by drag and drop, you will need to do it twice. First to add the Blender Lab repository, secondly to install the add-on itself. Blender says the drag-and-drop method also gets you update notifications. You can install from disk instead.
2. Open Edit → Preferences and check the Get Extensions section. MCP should appear among your installed extensions.
3. Go to Edit → Preferences → System and turn on Allow Online Access under Network.
4. Go to Edit → Preferences → Add-ons, expand the MCP entry and start the server.

In Claude Desktop:

1. Download the latest .mcpb bundle from the Blender MCP releases page. Ghosh double-clicked the file and confirmed when Claude asked whether to install the extension. Anthropic's documented method is Settings → Extensions → Advanced settings → Install Extension…, then select the .mcpb file. Use that route if double-clicking doesn't work. Menu names may differ depending on your Claude Desktop version and your organization's policies.
2. With Blender open, ask Claude to check whether it can connect to Blender.

What success looks like: Claude confirms the connection and can describe the scene you have open. In Claude Desktop, Anthropic says you can click the "+" button in the chat box and choose Connectors to see which MCP servers are connected and what tools they offer.

When it doesn't work:

• Extension won't install: Anthropic suggests updating Claude Desktop, downloading the file again in case it's corrupted, and checking your free disk space.
• Installed, but no tools show up: restart Claude Desktop, then check the extension's settings for missing fields.
• Workflow stops partway: Ghosh says his workflow failed when he minimized Blender, so keep it open. Blender doesn't document this requirement, so treat it as a user observation.
• Still stuck: Ghosh says you can give Claude the official setup guide and have it walk you through the process.

Section summary: four settings in Blender, one extension install in Claude and a connection test. Setup is easy. The hard part is deciding where to run it safely.

Why Not Just Use Computer Use?​

Ghosh addresses an obvious question. Claude's Computer Use feature can operate Blender through the screen, but he says it effectively takes over your PC. If you touch anything else, you interrupt it. It also burns more tokens because it keeps taking screenshots to see what's happening. MCP talks to Blender's API directly, so you can keep working while it runs. That's a real practical advantage. The trade-off is that it runs code, which is why the isolation advice above matters.

Is It "AI Slop"?​

Ghosh expects some readers to call this blasphemy. His position is that "slop" comes from shipping low-quality AI output to an audience, not from using AI to learn a tool. That's a fair point for a personal learning project. Being open about his clip's flaws also helps his case.

There is a counterpoint worth making. Watching Claude spend hours building scenes isn't the same as learning to model, rig or light them yourself. Ghosh says he feels more comfortable in Blender afterward, and that's believable. How much you actually learn probably depends on whether you study what the assistant built or just approve it. Blender itself describes its MCP exploration as coming from a project with a firm artist-centric policy on AI. Its own examples put the tool closer to an assistant that explains and inspects your work than one that replaces the artist.

The Bottom Line​

Blender's official MCP server really does lower the barrier for newcomers who want to try 3D work. Ghosh's four-hour robot short, flaws and all, shows what a patient beginner might manage. It may be the gentlest way into Blender yet, provided you're willing to revise repeatedly and to run it on a machine where AI-written code can't reach anything you'd miss.