using UnityMCP with ollama or local llm

[LivingRoomPC]

Hello, fantastic project, i was able to get it working with cursor in just a few minutes.

I was wondering if anyone has had success getting this to work with a local llm like gemma3 12b or qwen2.5 7b ?
I have searched up and down the internet, installed the dolphinMCP to try it out but with no luck. Was hoping someone could point me in the right direction, i found

ZundamonnoVRChatkaisetu /
unity-mcp-ollama

but i cant seem to get ollama to connect to the server
any help would be greatly appreciated

[LivingRoomPC]

Update: i got it to somewhat work with the continue vs code extension. was a bit tedious to setup but i was able to create a sphere and cube after a few very precise prompts. You really have to baby your prompt if you plan on using a local ai. or have a really good config file or else it wont be able to successfully run the tools. i am still testing, I think it would be cool to have the ai build player objects like a spaceship or a car using cubes and spheres.

I had to use a pretty specific prompt: "using the tools provided, create a cube in unity and name it playercube" if i didnt give it the name it would fail at running the manage_gameobject tool

here is my config.yaml for the continue local assistant using ollama llm. feel free to copy or give me tips on how to improve it.

name: Local Assistant
version: 1.0.0
schema: v1
models:
  - name: Autodetect
    provider: ollama
    model: AUTODETECT
context:
  - provider: code
  - provider: docs
  - provider: diff
  - provider: terminal
  - provider: problems
  - provider: folder
  - provider: codebase
mcpServers:
  - name: unityMCP
    command: uv
    args:
      - --directory
      - C:\Users\GTX3060\AppData\Local\Programs\UnityMCP\UnityMcpServer\src
      - run
      - server.py
rules:
  - >
    You are a Unity game assistant working with an external tool API. Your job is to manage GameObjects and other Unity assets through structured tool calls.

    Use the `manage_gameobject` tool for creating or modifying Unity primitives. Your output will be parsed and executed directly inside Unity's editor.
    ## Available Tools

    - `manage_editor`: Controls editor state and queries info.
    - `execute_menu_item`: Executes Unity Editor menu items.
    - `read_console`: Reads or clears Unity console messages.
    - `manage_scene`: Opens, closes, or modifies scenes.
    - `manage_gameobject`: Creates, modifies, or deletes GameObjects.
    - `manage_script`: Manages C# script files.
    - `manage_asset`: Loads or creates Unity assets and prefabs.

    ## Tool Example (Creating a Cube):
    json
    {
      "action": "create",
      "primitive_type": "Cube",
      "name": "MyCube",
      "position": [0, 0, 0]
    }
    
    To attach an existing script to a GameObject:
    {
      "action": "add_component",
      "gameobject_name": "Player",
      "component_type": "PlayerMovement"
    }

    Make sure the script exists first using `manage_script`. If it doesn’t, create it before attaching.
    
    ## General Guidelines
    - Prefer JSON-style output for tool calls.
    - Only include valid parameters for the tool being used.
    - Use camelCase for variable names.
    - If user intent is vague (e.g. "make a shape"), clarify it as best you can and choose a reasonable default (e.g. "Cube").
    - Always include a `position` when creating a GameObject.
    - If no name is provided, default to `"NewObject"` + object type.

    ## Scene Setup Tips
    - Ensure the scene includes a camera and a directional light.
    - Keep object names unique and meaningful.

[Efe-Oral]

I was looking for a way to send my prompt from Unity to the cursor LLM agent. Normally it's the other way around.
There has to be a LLM to decide and generate a structured request for which tool will the MCP server use and for that cursor's own agent works great, but I can't send a prompt from unity to cursor. Maybe with your way I don't have to use cursor's LLM after all, but you say that the prompts have to be perfect and precise, so it is less convenient compared to cursor's agent right?