Run Blender MCP Locally with Ollama (Free, No API Key)

If you want free, private, and offline AI-powered 3D modeling with Blender MCP, Ollama is the answer. This guide explains the setup — and an important caveat: Ollama does not natively speak the MCP protocol. The solution is to use an MCP-compatible client (like Cursor or VSCode with Cline) configured with Ollama as the model provider, alongside the Blender MCP server.

You get: local AI (Ollama) + Blender tools (MCP) = free AI 3D modeling with no API keys or cloud dependency.

How It Works

Ollama is an inference engine, not an MCP client. So you need a client that:

1. Supports MCP servers (to connect Blender MCP)
2. Can use Ollama as the model backend

Cursor and VSCode with the Cline extension both satisfy this. Cursor has built-in support for Ollama. Cline lets you point the AI to Ollama’s API. In both cases, you add Blender MCP as a separate MCP server.

Benefits of Local AI

• Free — No API costs, no usage limits
• Private — Your prompts and data never leave your machine
• Offline — Works without internet after models are downloaded
• Unlimited — No rate limits or quota caps

Limitations

• Slower — Local models are typically slower than cloud APIs
• Less capable — Models like Llama 3.1 are good but often less capable than Claude or GPT-4 for complex 3D workflows
• Hardware — You need a decent GPU (8GB+ VRAM recommended) for smooth performance
• Setup complexity — Two pieces to configure: Ollama + MCP client

Prerequisites

• Blender 3.0+ — Download from blender.org 
• Ollama — ollama.com 
• Cursor or VSCode (with Cline extension)
• uv package manager — For running Blender MCP
• GPU — 8GB+ VRAM recommended for local models

Step 1: Install Ollama and Pull a Model

curl -fsSL https://ollama.com/install.sh | sh

ollama pull llama3.1:8b

ollama pull llama3.1:8b

Step 2: Install Your MCP-Compatible Client

Option A: Cursor (Recommended)

Download Cursor . Cursor has native MCP support and can use Ollama as a model provider.

Option B: VSCode + Cline

Install VSCode  and the Cline extension . Cline supports MCP and can be configured to use Ollama.

Step 3: Configure the Client to Use Ollama

Cursor

1. Open Cursor Settings → Models
2. Add Ollama as a model provider (Cursor detects it when Ollama is running)
3. Select your pulled model (e.g., llama3.1:8b) for the Agent/Composer

VSCode + Cline

1. Open Cline settings
2. Set the API endpoint to your local Ollama instance (typically http://localhost:11434)
3. Choose your model from the Cline model selector

Step 4: Add the Blender MCP Server

Install uv if you haven’t:

curl -LsSf https://astral.sh/uv/install.sh | sh

Add the Blender MCP server to your client’s MCP config. The config is the same as in our other setup guides.

For Cursor — Go to Settings → MCP and add:

{
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["blender-mcp"]
    }
  }
}

{
  "mcpServers": {
    "blender": {
      "command": "cmd",
      "args": ["/c", "uvx", "blender-mcp"]
    }
  }
}

For VSCode + Cline — Add the same config to Cline’s MCP settings (location varies by Cline version; typically in workspace or user settings).

Step 5: Install the Blender Addon and Connect

Step 6: Test Your Setup

With Blender connected and your client configured:

1. Open the Agent/Composer (Cursor) or Cline chat in VSCode
2. Ensure the model is set to your Ollama model
3. Verify Blender MCP appears in the tool list
4. Try: “Create a red cube in Blender” or “Add a UV sphere at the origin”

Local models may need clearer, shorter prompts. If a complex request fails, break it into smaller steps.

Troubleshooting

Ollama model doesn’t use Blender tools

• Confirm the model supports function/tool calling (Llama 3.1, Mistral, and CodeLlama do)
• Run a simpler prompt first to verify the MCP connection
• Some smaller models struggle with multi-tool workflows — try a larger model

Blender MCP shows disconnected

• Same fixes as other guides: verify uv, Blender addon enabled, Connect clicked
• Restart the client after adding MCP config
• Only one MCP client should connect to Blender at a time

Slow or unresponsive

• Local inference is slower than cloud APIs
• Use a smaller model (llama3.1:8b) for faster iteration
• Ensure your GPU is being used (nvidia-smi on Linux to check)

Out of memory

• Use a smaller model or reduce context length
• Close other GPU-heavy applications

For a technical deep dive, see our MCP Server architecture guide.