Skip to main content

Blender MCP

TECHNIQUE
📖6 min read
Control Blender With AI Using Python and MCP by Real Python
1 / 2

Introduction​

The Model Context Protocol (MCP) is an open standard for connecting Large Language Models to external tools and data. An MCP server for Blender gives an LLM access to your running Blender session: it can query scene data, rename and organize objects, run checks, and execute modeling operations through natural language.

The highest value for investigation work is the read and organize side: scene analysis, naming conventions, inventory queries, checklists, and reports. That output is reviewable and belongs in a case record.

caution

You are giving an LLM access to a running Blender session. The server executes what the model decides to call, as your user, without guards. Scene metadata leaves your machine when you use a cloud client. Works well for simple and structured tasks, not a replacement for judgment on sensitive operations.

This page covers the technique: choosing a server, the two setup paths (cloud and local), and how to document the work. The companion Blender MCP addon page documents each featured server in detail.

Key Takeaways​

Two Setup Paths

Cloud clients (Claude) are the easiest start. Local LLMs (Ollama, llama.cpp) keep scene data on your machine. The addon and server setup is identical either way.

Organizational Value

Naming, inventory, checklists, and scene reports are the defensible use cases. The output is diff-able in the Outliner and exportable to the case record.

Verification Discipline

LLM output is not deterministic and code execution is not sandboxed. Verify every change, diff before and after, and keep a backup copy before each session.

What is the Model Context Protocol?​

MCP standardizes how applications expose tools and data to LLMs. A server defines:

  • Tool definitions. What operations the model can perform, each with a schema.
  • Resource access. What data the model can read.
  • Two-way communication. The model calls a tool, the server executes it in Blender's Python API and returns the result.

The model never touches your .blend file directly. It asks, and the server does the work in your running session.

Choosing a Server​

Three criteria, applied strictly (full comparison on the addon page):

  1. Organizational value. Analysis, renaming, inventory, checklists, not just prompt-to-geometry.
  2. Privacy and security posture. No default-on telemetry, clear code-execution warnings, a path to local models.
  3. Maintenance. Recent commits and active issue handling, not star count.
ServerUse it for
blender_mcp (official)Scene analysis, renaming conventions, documenting setups, checklists
blender-ai-mcp (PatrykIti)Verification-first work: deterministic measurement and assertion tools
MCPBlender / blender-mcpModeling and asset pipelines. The most popular server, but telemetry is on by default, so disable it for real work

Workflow: Cloud Setup​

1

Install uv and the MCP server

uv is the package manager that runs the server. macOS: brew install uv, Linux: curl -LsSf https://astral.sh/uv/install.sh | sh, Windows: powershell -Command "irm https://astral.sh/uv/install.ps1 | iex".

2

Install the Blender addon

For MCPBlender: uvx blender-mcp install-addon. For the official server: install the add-on zip from the Blender Lab releases. Then enable it in Edit, Preferences, Add-ons.

3

Register the server in your client

Claude Desktop: Settings, Developer, Edit Config. VS Code: the MCP panel or a .mcp.json file. Start with a read-only query before you ask for changes.

Example server config for MCPBlender:

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

With a cloud client, scene metadata (object names, positions, materials, and queries) is sent to the provider. The 3D geometry stays local. MCPBlender additionally runs a default-on telemetry pipeline that, with consent, collects prompts, generated code, scene metadata, screenshots, and action trajectories. The official and PatrykIti servers do not. For sensitive investigations, use the local path below.

Workflow: Local Setup (Ollama)​

1

Install Ollama and pull a model

From ollama.ai, then ollama pull qwen3-coder (7B minimum, 14B+ for complex tasks).

2

Start the server

ollama serve (listens on http://localhost:11434).

3

Point your MCP client at it

Example for MCPBlender with Ollama:

{
    "mcpServers": {
        "blender": {
            "command": "uvx",
            "args": ["blender-mcp"],
            "env": { "BLENDER_MCP_LLM": "ollama", "BLENDER_MCP_LLM_MODEL": "qwen3-coder" }
        }
    }
}

Hardware expectations: 8GB+ GPU VRAM for comfortable use, CPU-only works but expect 5 to 30 second responses, 16GB+ system RAM for large scenes.

What MCP Can Do for You​

Scene organization and analysis​

The defensible use cases, with the prompt patterns that work:

  • Systematic renaming. "Rename all objects starting with measure_ to MEASUREMENT_ followed by their index."
  • Collection structure. "Create collections for cameras, lights, and geometry. Move objects to their collections by type."
  • Geometric analysis. "Find all objects below height 0 and report their locations and bounding box sizes."
  • Scene reports. "Generate a structured report of the scene: object count, collection tree, material count, absolute paths."

Scene manipulation​

  • Create, delete, and modify objects, transforms, and materials
  • Add and modify modifiers, parent and unparent hierarchies
  • Tag objects with custom properties
  • Import assets and HDRIs (MCPBlender adds Poly Haven, Sketchfab, and Poly Pizza integrations)

In an Investigation​

To keep your work aligned with the Berkeley Protocol and the Guide for Judges, document MCP use in three places.

Method​

  • Objective. What you automated (for example: scene organization, data querying, checklist verification).
  • Tool versions. The MCP server (name + version), the client and model, the Blender version.
  • Inputs. Paste the exact prompts you used for each operation.

Decision Log​

  • Settings. Model (name/version), temperature, system prompt, server configuration.
  • Assumptions. Acknowledge that the LLM may misinterpret instructions or generate incorrect code.
  • Why AI-assisted. Record why an automated step was used for this operation (for example: bulk renaming of 400 imported objects to the case naming convention).

Verification​

  • Inspect every AI-executed change (names, locations, properties) in the Outliner before saving.
  • Compare the object list before and after each session.
  • Validate results against a manual check or a deterministic measurement tool where the server provides one.

Example methodology statement:

Scene organization was automated using blender_mcp (official, v1.0.0) with a local qwen3-coder model via Ollama. The instruction "Rename all objects in collection X to the case naming convention" was executed. All resulting name changes were manually verified against the project naming convention, and the object list was diffed before and after. No geometric modifications were performed by the AI.

Limitations​

  • The server executes LLM-generated code without guards. The official documentation recommends a virtual machine for untrusted work.
  • LLMs generate incorrect or destructive code. Undo is your last line of defense, and it is not reliable. Save before every session.
  • LLM analysis is not deterministic. It misses things. Verify in the Outliner.
  • Cloud clients send scene metadata to the provider. MCPBlender adds a default-on telemetry pipeline.
  • Local models are slower and weaker than cloud models.
  • Complex geometric operations are often beyond current text-based models.
  • Run only one MCP client instance per Blender session.
See Also

The Blender MCP addon page covers the featured servers, their setup, and organizational example prompts in depth. For manual techniques, see the Scene Management guide.

Summary​

MCP gives an LLM a working interface with your Blender session. You pick a server, wire one client, run the organizational jobs, and verify every change with diffs before the result enters a case record.

Resources​

Sources​

Tags:
Last updated on by blender-visual-investigation