Blender AI Control via MCP
/mcp-for-blender
mcp-for-blender bridges AI assistants and Blender through nine focused tools exposed via the Model Context Protocol, acting as the essential pipe when AI cannot reach Blender natively. Setup is surprisingly smooth: one shell command handles everything, and the addon even backs up config files first. Results depend on your LLM's coding skillβsimple tasks work reliably, complex ones may need a few retries. Worth trying if you prefer staying in a chat window rather than hunting through Blender's UI.
Connecting AI Assistants to a 3D Viewport
MCP for Blender connects AI assistants like Claude to Blender through the Model Context Protocol, letting you control a 3D viewport with natural language. The AI generates and executes Blender Python scripts directly, so you describe what you want and Blender builds it. It runs as a Blender addon paired with an MCP server that exposes nine tools β object creation, material control, camera views, asset search, and AI model generation β without requiring you to memorize a single API call.
Nine Tools the AI Can Use
- βNine MCP tools: `execute_blender_code`, `look` (multi-angle and wireframe views), `get_scene_info` (compact scene summary), `generate_3d` (Tripo, Hunyuan3D, Hyper3D Rodin), `search_assets` / `import_asset` (Poly Haven, Sketchfab, Poly Pizza)
- βTwo-way communication via a socket server embedded in the Blender addon β AI executes code and receives visual feedback
- βSafe mode: blocks scripts that read/write files, run subprocesses, or persist code after execution; returns blocked scripts with a reason so the AI can revise
- βMulti-client support across Claude Desktop, Claude Code, Codex, Cursor, VS Code, OpenCode, and Antigravity
- βDocker deployment: the MCP server runs in a container while Blender stays on the host, with automatic host detection on macOS/Windows and host networking on Linux
- βPoly Haven asset search, Sketchfab model import, and low-poly model sourcing from Poly Pizza β all surfaced as thumbnail pickers in Codex
One-Command Setup Across macOS, Windows, and Linux
The one-line installer works across macOS, Linux, and Windows (PowerShell) and handles the entire setup in one pass: installs uv if missing, detects Claude Desktop and other supported clients, adds the MCP server entry with a backup of each config file, and installs and enables the Blender addon. Restart your AI client fully (quit from the system tray on Windows), open Blender, and the addon activates automatically β click Start MCP Server in the sidebar if needed. Manual setup is three steps: install uv, add the MCP server entry to your client's config, run `uvx mcp-for-blender install-addon` and enable it in Blender's preferences.
Strengths
- βKeeps you in the AI chat window instead of tabbing between a prompt and Blender's UI
- βThe installer is thorough β it detects clients, backs up configs, and covers macOS, Linux, and Windows
- βNine tools cover the full workflow: code execution, visual feedback, scene queries, and asset/AI model import
- βSafe mode adds a security layer for scripts without blocking normal Blender operations
- βDocker option isolates the server, making it easy to reset state or run on a different machine
What MCP for Blender Cannot Do (Yet)
- β³The socket server has no authentication by default β anyone who reaches the port can run Python in Blender. Keep it on localhost and prefer an SSH tunnel for remote access rather than exposing the port directly
- β³Script quality depends entirely on your LLM's ability to write correct Blender Python β complex modeling tasks may need multiple retries or manual fixes
- β³Only one AI client can run the MCP server at a time; the README explicitly warns against running both Cursor and Claude Desktop simultaneously
- β³Safe mode blocks file I/O and network access in scripts, which can interfere with workflows that legitimately need those operations from within Blender
- β³Setup friction on Windows β PATH configuration for uv is non-obvious, and GUI-launched clients may not see uvx without the full path or a wrapper
Alternatives for AI-Assisted 3D Workflows
Who Should Use MCP for Blender?
3D artists who want to describe a scene and have it appear in Blender without touching the UI. Technical artists who need to automate repetitive tasks like material or collection cleanup. Indie developers who want to prototype 3D layouts quickly before committing to detailed work. Studios exploring AI-assisted scene composition and lighting without abandoning their existing Blender workflows.
Frequently Asked Questions
MCP for Blender is free and open source under the MIT license. Premium AI 3D model generation (Tripo, Hunyuan3D, Hyper3D Rodin) requires a paid subscription, but the core plugin and all standard features are free.
Claude Desktop, Claude Code, Codex (CLI, desktop app, and IDE extension), Cursor, VS Code, OpenCode, and Antigravity. Only one can run the MCP server at a time.
Blender 3.0 or newer. The addon creates a socket server inside Blender and needs the N-panel sidebar, which has been standard since 2.93.
No β only one MCP server instance is supported. The README explicitly warns against running both Cursor and Claude Desktop simultaneously. Multiple Blender instances on different ports are supported, but each needs its own client config.
Yes. The addon runs the socket server that bridges your AI client to Blender. If you close Blender or disable the addon, the connection drops. The N-panel sidebar shows the server status and lets you restart it.
Set BLENDER_MCP_SAFE_MODE=1 and the addon validates scripts before execution. It blocks file I/O, subprocess calls, network access, and code that persists after the script ends. Normal Blender operations (modeling, materials, rendering) are unaffected. Blocked scripts are returned to the AI with an explanation, giving it a chance to revise before trying again.
The problem it solves
Creating or iterating on 3D scenes requires constant context-switching: look up a Python command, open the script editor, run it, check the result, tweak, repeat. Even a straightforward material change breaks the creative flow. There was no way to drive Blender from an AI assistant in natural language without manually translating prompts into scripts.
Best use cases
- β’Rapid 3D scene prototyping: describe a room layout or an object arrangement and watch it appear in Blender
- β’Automating repetitive Blender tasks β material swaps, collection organization, lighting adjustments β by prompting instead of scripting
- β’AI-assisted workflow pipelines: use Tripo or Hunyuan3D to generate a base mesh, then hand it off to Blender for refinement
- β’Quick layout exploration before committing to detailed modeling, especially useful under tight deadlines
Related repositories
Is mcp-for-blender worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of mcp-for-blender.