Guides
Blender MCP with a local LLM: run an AI assistant for Blender offline
Blender MCP with a local LLM: the official llama.cpp setup, Ollama options, hardware, model choice and tips for reliable tool calls.
3 min read

Quick answer
Can Blender MCP run with a local LLM?
Yes. The official Blender MCP project documents a local setup: llama.cpp (build 8218 or newer) serves a model, the blender-mcp server provides the tools, and Blender 5.1 runs the add-on. Nothing leaves your machine, but you need a model that handles tool calls well and enough RAM or VRAM.
Why run Blender MCP locally?
A local LLM means no subscription, no usage cap, and no scene data leaving your machine. That matters for studio work under NDA and for people who use Blender MCP for hours at a time. The trade-off is speed and reliability: local models are slower, and smaller ones make more mistakes with tool calls.
The official local setup (llama.cpp)
Blender Lab's MCP project includes a dedicated guide, readme_local_llm.rst, in its repository. Its prerequisites are the repo itself, llama.cpp version 8218 or newer, and Blender 5.1. The components:
- llama.cpp runs the model.
- chat_client is a simple text client that connects llama.cpp to the MCP server.
- blender_mcp is the MCP server with the tools.
- blender_mcp_addon connects Blender to the server.
The steps, summarised from that guide:
- Create a Python virtual environment and install the server:
pip install -r mcp/requirements.txt, thenpip install -e mcp. - Build and install the Blender extension:
blender -c extension build --source-dir ./addon/blender_mcp_addon, thenblender -c extension install-filewith the built zip. - Download and run a model with
llama-server --jinjaand a GGUF from Hugging Face. Checkhttp://127.0.0.1:8080/responds. - Run the chat client against it, or use llama.cpp's web UI, which the guide says gives a nicer experience with progress feedback.
- In Blender, enable Online Access and make sure the add-on's MCP Bridge Server is running in its preferences.
Using Ollama or LM Studio instead
The Blender MCP server only sees MCP tool calls; it doesn't know what model made them. So any MCP client that can run local models, with Ollama or LM Studio as the backend, can use Blender MCP. Register the server command (blender-mcp, or uvx mcp-for-blender for the community server) in that client the same way you would for Cursor. See Blender MCP with Codex, Cursor and Gemini.

Picking a local model for Blender
What matters most is tool calling, not raw benchmark scores. A model that writes beautiful Python but formats tool calls wrongly is useless here. When testing:
- Start with a read-only task: "List objects and their materials."
- Then a small write task: "Add a cube at the origin, scale it to 2 m."
- Then a multi-step one: "Make a table from cubes: top 1.2 × 0.8 m, four legs 75 cm tall."
If a model fails the second test often, try a larger one or a different family. The official guide's example uses a large Qwen GGUF; use it as a hint, not a requirement.
Hardware
Large local models need lots of memory: GPU VRAM on a PC, unified memory on Apple silicon. If generation is unbearably slow, use a smaller quantisation or a smaller model, and keep requests short. The same hardware thinking applies to running 3D generators locally; see GPU for AI 3D generation.
Keep it safe even offline
Local doesn't mean safe. The model still runs Python inside Blender. Save often, and follow the official project's advice to use a VM or isolated system for anything important.
Local assistant, generated models
A local LLM drives Blender well but doesn't make detailed meshes. Generate those with a 3D model, locally (local AI 3D model generator) or in Vi3W in the browser with no GPU, then let your local assistant arrange them. Try Vi3W.
Frequently asked questions
blender mcp with ollama
The official project's documented path is llama.cpp. Ollama can work through any MCP client that supports Ollama models, because the Blender server doesn't care which model is calling it. Pick a model with reliable tool calling.
blender mcp with local llm
Follow the official readme_local_llm guide: install the MCP server, build and install the Blender extension, run llama-server, then use the included chat client or llama.cpp's web UI.
blender mcp token usage
Screenshots and full scene summaries are the expensive calls. Ask for specific objects instead of whole-scene dumps, and take screenshots only when you need to check a result. Locally, tokens cost time instead of money.
Last reviewed

