Blender Automation

Blender Automation via MCP

Connect a running Blender instance to Hermes via the Model Context Protocol, enabling natural-language 3D modeling, scene manipulation, material editing, and asset generation.

When to Use

Use this whenever the user wants to:

• Create, modify, or delete 3D objects by describing them in plain language
• Inspect and analyze Blender scenes (poly counts, materials, modifier stacks)
• Apply batch operations across many objects (rename, recolor, reposition)
• Generate Blender add-ons, operators, and panels without writing Python
• Set up rendering pipelines, lighting, and cameras conversationally
• Download and apply Poly Haven HDRIs / textures programmatically
• Generate AI 3D models via Hyper3D Rodin integration
• Advise on the broader AI+Blender landscape — texture generation tools, 3D model generation (Hyper3D, Meshy, Tripo, Hunyuan3D), AI animation (AnimateDiff, Rokoko mocap), and concept art pipelines (ControlNet). See references/ai-blender-ecosystem.md for a condensed reference.
• Assess hardware viability — proactively note Intel Mac limits (no GPU for local diffusion, cloud workarounds needed for texture gen and AI animation, Eevee preferred over Cycles for rendering speed)

Architecture Overview

Hermes Agent ──stdio MCP──→ uvx blender-mcp ──TCP socket 9876──→ Blender Addon
                                                                      │
                                                                  Blender (bpy API)

• MCP Server (uvx blender-mcp): Bridges Hermes ↔ Blender. Hermes's native MCP client spawns this as a subprocess.
• Blender Addon: Runs inside Blender as a socket server on port 9876. Receives commands from the MCP server and executes them via bpy.
• Tool Naming: All tools appear as mcp_blender_* in Hermes (e.g., mcp_blender_get_scene_info, mcp_blender_create_object).

Two Implementations

1. Community: ahujasid/blender-mcp (Recommended for now)

• Repo: github.com/ahujasid/blender-mcp — MIT, 22.6k stars (Jun 2026)
• Requirements: Blender 3.0+, Python 3.10+, uv
• Features: Scene inspection, object CRUD, materials, arbitrary bpy execution, viewport screenshots, Poly Haven assets, Hyper3D AI models, Sketchfab search
• Install: uvx blender-mcp (MCP server), addon from repo (Blender addon)
• Ecosystem: 107+ repos on GitHub matching "blender MCP server". MCP is the de facto standard protocol — ecosystem has converged. Multi-LLM support (Claude, ChatGPT, Gemini, Ollama) is table stakes.

2. Official: blender.org MCP Server

• Source: blender.org/lab/mcp-server
• Requirements: Blender 5.1+ (not yet released)
• Built by Blender dev team as part of their lab program
• Pair with any MCP client (Llama.cpp, Claude, etc.)

3. Notable Alternatives

4. Emerging Pattern: Agent Skills for Blender

• cc-blender-skill — 10 chain-loadable skills (modeling, materials, lighting, rendering, animation, export, wireframe-to-3d). Validated on 6 scene classes for Blender 5.x
• blendops — AI-native workflow pack: Blender MCP + Claude + Three.js/React Three Fiber

5. Moonlake AI — Computer-Use Agent (Not an Addon)

• Approach: Operates Blender via screen interaction (like Anthropic's computer use), NOT a Blender addon/plugin
• Architecture: Two-model system — multimodal reasoning model (causality, physics, persistence) + RIE neural rendering model (style transfer)
• Philosophy: "Structure, not scale" — language + symbolic representations are 5 orders of magnitude more data-efficient than pure pixel/video diffusion
• Capability: Iterative, long-horizon refinement. Builds articulated assets, physics-validated scenes, complex environments. Can learn from expert demos.
• Status: Beta since Feb 2026, $15/mo. Partners with $8T+ combined market cap.
• GitHub: github.com/MoonlakeAI — content-agents (VLM-powered USD workflows), Nova3D (Prompt-to-Code targeting Blender Python API)
• UE5 relevance: Explicitly mentions Unreal Engine compatibility in asset management pipeline

Setup

Prerequisites

# Ensure uv is installed
which uv || brew install uv   # macOS
# or: https://docs.astral.sh/uv/getting-started/installation/

Step 1: Install Blender

• Download from blender.org or brew install --cask blender
• Requires 3.0+ for ahujasid/blender-mcp

Step 2: Install the Blender Addon

1. Download addon.py from the ahujasid/blender-mcp repo root
2. Open Blender → Edit > Preferences > Add-ons
3. Click Install… and select addon.py
4. Enable "Interface: Blender MCP"
5. In the 3D View sidebar (press N), switch to the BlenderMCP tab
6. (Optional) Enable Poly Haven checkbox for HDRI/texture assets
7. Click "Connect to Claude" — this starts the socket server on port 9876

Step 3: Configure Hermes

Add to your Hermes config (~/.hermes/profiles/senna/config.yaml for profile-isolated setups):

mcp_servers:
  blender:
    command: "uvx"
    args: ["blender-mcp"]

Then restart Hermes. Tools are auto-discovered with the mcp_blender_* prefix.

⚠️ Only run one MCP client at a time. If another app (Cursor, Claude Desktop) also has blender-mcp configured, stop it first — only one uvx process can connect to the Blender addon.

Step 4: Verify

source ~/.hermes/hermes-agent/venv/bin/activate
hermes mcp list
# → blender should appear with ✓ enabled
hermes mcp test blender
# → Should show ✓ Connected and ✓ Tools discovered: N

Using Blender Tools from Hermes

Once connected, you can ask Hermes to do things like:

• "Create a low-poly dungeon scene with a dragon guarding treasure"
• "Inspect the current scene and list the top 5 highest-poly objects"
• "Make this car red and metallic, then set up studio lighting"
• "Rename all objects to use consistent naming: OBJ-material-number"
• "Take a viewport screenshot and analyze the topology"
• "Generate a 3D model of a garden gnome via Hyper3D"
• "Create a sphere above the cube, rotate the cube 45°, and point the camera at the scene"
• "Apply subdivision surface to all selected objects with 2 levels"

Environment Variables

Set these in the env block of the MCP config if customizing.

Critical Architecture Lessons (from 3D-Agent)

Research on multi-agent Blender systems revealed several hard-won lessons:

1. Vision Feedback Loop is Mandatory

After each operation, take a viewport screenshot and have the agent verify it before the next step. Without this, geometry degenerates into "soup" after 3-4 operations. The agent should also learn to change viewport perspective when needed.

2. Separate Reasoning from Code Generation

One model decides what to do (reasoning/planning), another generates how in bpy (code execution). This avoids the single-model drift problem.

3. DSPy for Prompt Optimization

Train system prompts against Blender API docs and expected outputs using DSPy — far more consistent than hand-tuning.

4. Math-Heavy Operations Are Weak

Precise geometric calculations (lattice structures, curve math, symmetry) may need a dedicated pass. Not yet reliable from raw LLM prompting.

5. bpy API Docs Are Insufficient Alone

RAG over docs + examples gets only ~50% coverage. Supplement with example-heavy prompts and iterative refinement.

6. TCP Socket > HTTP for Blender Communication

The addon uses a socket directly rather than HTTP. This matters for the 30+ back-and-forth operations per task — responsiveness is critical.

Capabilities (ahujasid/blender-mcp v1.5.5)

Troubleshooting

References

• references/blender-mcp-research.md — Detailed session notes: provider comparison, architecture diagrams, community resources.
• references/ai-blender-ecosystem.md — Broader AI+Blender landscape reference: texture gen tools, 3D model generators, AI animation, concept art pipelines, Intel Mac hardware reality table.