blender-mcp-ultra

The most complete MCP server for Blender — control Blender with natural language from any AI agent: Claude Code, Claude Desktop, Cursor, opencode, Windsurf, Antigravity, LM Studio, Ollama and any Model Context Protocol (MCP) client.

239 tools · Windows / macOS / Linux · headless-ready · production-grade security

Scene above: 100% created and rendered by an AI agent through this MCP server — no manual modeling.

Gallery — what agents build with it

| PBR materials | Geometry Nodes | Physics | |---|---|---| | | | | | material_pbr + 3-point lighting | geonodes scatter + noise terrain | rigid body + baked simulation |

Every image above was generated end-to-end through MCP tools — scene setup, materials, lighting, camera framing, simulation and final Cycles render.

---

What is it?

blender-mcp-ultra connects Blender (the free open-source 3D suite) to LLM agents through the Model Context Protocol. Ask your AI assistant to model, texture, light, animate, simulate and render 3D scenes — it executes through 248 typed tools instead of fragile screenshot-guessing.

You:     "Create a wooden table with a coffee mug on it and render it in Cycles"
Claude:  → object.create, mesh ops, material PBR wood, light.three_point,
          camera.set_framing, render.render → escena.png

Highlights

• 🧰 248 MCP tools (7 base + 241 registry, incl. built-in workflow guidance) — modeling, materials, shader nodes, geometry nodes, lighting, cameras, animation, physics, compositor, UV, rigging, 3D printing, export (FBX/OBJ/glTF/STL), batch, VLM visual feedback, multi-agent collab, scene planner
• 🖥️ Any MCP client — stdio (Claude Code/Desktop, Cursor, opencode, VS Code, Windsurf), SSE/HTTP (Antigravity, remote), REST
• 🧠 VLM feedback loop — the agent sees its renders (EEVEE headless capture) and iterates
• 🔒 Security-first — AST code blocklist validated in gateway and addon, mandatory auth token (auto-generated), localhost-only binding, rate limiting, 0 secrets in history (gitleaks)
• 🤖 Multi-provider — OpenAI, Anthropic, Google, DeepSeek, Ollama (local)
• 📦 Asset integrations — PolyHaven, Sketchfab, AmbientCG, Hyper3D
• 🚀 Headless CI — real E2E tests against Blender in GitHub Actions (Linux + Windows + macOS)
• 🧪 479 tests passed · ruff · bandit · pip-audit clean

Quick Start

1. Install the Blender addon

Download blender_mcp_ultra.zip → Blender → Edit > Preferences > Add-ons > Install… → enable blender-mcp-ultra → click Connect in the N-panel.

Or install from this repo (Blender 4.2+/5.x):

git clone https://github.com/carlosh7/blender-mcp.git
cd blender-mcp
pip install -e .

2. Connect your AI client

Install the gateway (provides the blender-mcp-server command). Startup order doesn't matter: all tools register even if Blender is closed and work as soon as you open it.

pip install blender-mcp-ultra

<details open> <summary><b>Claude Code</b></summary>

claude mcp add blender -- blender-mcp-server
# o, desde un checkout del repo:
claude mcp add blender -- python /ruta/a/blender-mcp/mcp_server.py

More: docs/clients/claude-code.md

</details> <details> <summary><b>Claude Desktop / Cursor / Windsurf / VS Code</b> (mcp.json)</summary>

{
  "mcpServers": {
    "blender": {
      "command": "blender-mcp-server"
    }
  }
}

More: docs/clients/claude-desktop.md · cursor.md · vscode.md

</details> <details> <summary><b>opencode</b></summary>

python mcp_server.py            # el addon auto-escribe ~/.config/opencode/mcp.json

More: docs/clients/opencode.md

</details> <details> <summary><b>Antigravity / clientes HTTP / remoto</b></summary>

python mcp_server.py --sse      # SSE en :9879 (MCP_SSE_HOST para exponer)
# REST local: http://localhost:9877 (mismo token que el socket: ~/.config/blender-mcp/socket_token)
# Docker:
docker build -t blender-mcp-gateway . && docker run -p 9879:9879 blender-mcp-gateway

More: docs/clients/antigravity.md · docs/remote.md

</details>

3. Ask your agent

"Modela una silla de madera con biselado y material PBR, ilumina en 3 puntos y renderiza"

Tools (248)

| Category | Examples | |---|---| | Scene | scene.get_info, scene.create, scene.render_settings, scene.query | | Objects & Mesh | object.create/transform/join, mesh.get_topology, mesh.bevel_edges, mesh.extrude_faces | | Materials & PBR | material_create, material_pbr (wood/metal/fabric/leather/stone/glass…), material_assign | | Shader / Geo Nodes | shader.add_node/connect_nodes, geonodes.scatter/array/create_group | | Lights & Camera | light.three_point, light_create, camera.track_to, camera.set_framing | | Animation | animation.set_keyframe, animation_get_fcurves, NLA, drivers, shape keys | | Physics | rigid body, cloth, soft body, particles, force fields, constraints, baking | | Render | Cycles/EEVEE settings, render_render, render_viewport, filmic, jobs con progreso | | I/O & Export | FBX, OBJ, glTF, STL + export_for_target (game/web/print/film), LODs, colisiones | | UV & Texture | smart UV project, unwrap, bake, textures | | Rigging | armatures, bones, constraints, vertex groups, auto weights | | 3D Printing | manifold/watertight checks, scale to mm, dimensions | | Batch & Perf | batch rename/materials/modifiers, turntable, LODs, memory report | | VLM Feedback | vlm_capture, vlm_analyze, composition & lighting checks | | Collab & Planner | locks por objeto, mensajes entre agentes, tareas, plan_create/execute | | Versioning & Docs | vc_snapshot/restore, docs_scene, eventos poll_events |

Full reference generated from the live registry: docs_scene tool · skills/recetas en docs/skills/.

Security

| Layer | Mechanism | |---|---| | Transport | localhost-only sockets + shared auth token (auto-generated; X-API-Token on HTTP) | | Code execution | AST blocklist (code_guard, ~80 patterns) in gateway and addon + 10s timeout + auto-undo | | Input | rate limit (30 req/min) and 1 MB body cap on the HTTP API | | Audit | structured JSON logs with rotation | | Supply chain | gitleaks (full history) + bandit + pip-audit in CI |

Run trusted agents only — executing Blender Python on the agent's behalf is the product.

Architecture

AI Agent (Claude/Cursor/…) ──MCP stdio/SSE──▶ mcp_server.py (gateway, 248 tools)
                                                  │ socket :9876 (localhost + token)
                                                  ▼
                                    Blender + addon blender-mcp-ultra
                                    (handlers, validators, PBR, physics…)

mcp_ultra/  # hexagonal: core / adapters / infrastructure / presentation / tools
addon/      # Blender addon: socket server, handlers, PBR factory, anti-blockout
docs/       # client guides, skills recipes, tutorials
tests/      # 479 tests (unit + integration + e2e headless)

Testing

pytest tests/unit -q          # sin Blender
pytest -q                     # completo (requiere Blender con el addon conectado)

CI ejecuta la suite completa contra Blender headless en cada push (Linux, Windows, macOS).

FAQ

Does it work with ChatGPT / Gemini? Any client that speaks MCP works. For providers without MCP support, the bundled REST API (:9877) accepts plain HTTP.

Does Blender need to be open? Yes for interactive work (it drives the real Blender). CI/headless render jobs run with blender -b — see docs/skills/headless_ci.md.

Is it safe? The agent can execute Blender Python — that's the product. Execution is sandboxed, validated by AST, rate-limited and bound to localhost with token auth. Run trusted agents only.

Blender versions? 4.2+ (LTS) and 5.x. Python 3.10+ for the gateway.

Contributing

PRs welcome — read CONTRIBUTING.md. Docs: GUIDE.md · ARCHITECTURE.md · CHANGELOG.md

License

MIT — see LICENSE.