Blender MCP Tutorial: Control Blender from Claude and Cursor Agents
Step-by-step guide to mcp-for-blender and Blender MCP addons—install the MCP server, connect Claude Desktop or Cursor, and run safe 3D workflows without leaving Blender.
Posted by

Why Blender MCP matters in 2026
3D artists and technical directors want agents to move objects, apply modifiers, and render previews—but exporting to a separate web UI breaks muscle memory. Open-source projects such as mcp-for-blender (often called Blender MCP on social posts) expose Blender operations through the Model Context Protocol so Claude, Cursor, or a custom harness can drive the scene graph locally.
This tutorial walks through a pragmatic setup: install the addon, run the MCP bridge, connect a client, and execute guarded tool calls. It is editorial guidance—not affiliated with the addon authors—verify versions on GitHub before production use.
Prerequisites
- Blender 3.x or 4.x installed locally
- Python environment matching the addon README
- Claude Desktop, Cursor, or another MCP host
- Back up .blend files before granting agent write access
1. Install the Blender MCP addon
Clone or download the repository from GitHub (ahujasid/mcp-for-blender). Follow the project README to copy the addon into Blender's scripts path or install via the Preferences → Add-ons → Install flow. Enable the addon and note any required companion Python package the README lists for the MCP stdio or socket bridge.
Community forks may rename the addon; search for "blender-ai-mcp-addon" if your distro uses a different slug. Stick to repos with recent commits and clear license terms.
2. Start the MCP server bridge
Most Blender MCP stacks run a small Python process that speaks MCP on stdio while forwarding commands to Blender's Python API. Start Blender with the addon enabled, then launch the bridge script exactly as documented—often a uv run or python -m command with a config path.
# Example shape — use the repo's exact command
cd mcp-for-blender
python -m blender_mcp.serverKeep Blender open while the bridge runs; closing Blender will drop active tool sessions.
3. Register the server in Claude Desktop or Cursor
Add an MCP server entry pointing at the bridge command and working directory. In Claude Desktop, edit claude_desktop_config.json; in Cursor, use MCP settings with the same command vector. Restart the host after saving.
Confirm the tool list appears—typical tools manipulate objects, materials, or viewport renders. If tools are missing, check Blender's system console for Python tracebacks.
4. Prompt patterns that work
- Inspect before mutate: ask the agent to list objects and materials before destructive deletes.
- Small steps: one modifier or one lighting change per turn—easier to undo than a monolithic "rebuild the scene" request.
- Render checkpoints: request a low-sample viewport or preview render after major changes.
- File hygiene: save versioned .blend files (project_v001.blend) before agent sessions.
5. Safety: pair with policy and secret hygiene
Blender MCP can execute powerful Python inside your project. For teams, combine local setups with an agent tool-call guard mindset—audit tool calls before enforcing denylists on dangerous operations—and LocalGuard if assistants might read render farm credentials from env files.
For broader agent memory across repos, compare shared agent context workspace patterns so 3D prompts do not live only in chat history.
6. Alternatives and next steps
ComfyUI 3D nodes and proprietary DCC plugins compete for agent attention; Blender MCP wins when you must stay inside Blender's native toolset. Watch GitHub stars and issue velocity on mcp-for-blender—community buzz in late 2026 reflected fast iteration, but treat star counts as orientation, not quality scores.
Browse related MCP infrastructure on AIXList: NitroStack for hosted MCP deploys, OpenScreen for agent video, and the Mobbin MCP hub for design-reference agents.