Image Vision MCP

MCP Server that provides descriptions of images.

Requirements

Ollama

You must have Ollama set up, and exposing its LLMs via its API / webserver.

By default, the MCP uses the llava:34b LLM, which you need to have installed and on the same computer that Ollama is running on.

You can install the model via:

ollama run llava:34b

You can also specify a different model installed via Ollama via the --model argument in Claude config.

Installation

In order to install the MCP into Claude, add the following to the claude_desktop_config.json file.

    "image-vision": {
      "command": "npx",
      "args": [
        "-y",
        "node",
        "C:/Users/FOO/src/image-vision-mcp/src/image-vision-mcp.js",
        "--permitted",
        "C:/Users/FOO/mcp",
        "--host",
        "http://192.168.1.238:11434"
      ]
    }

Arguments

• --permitted (required) The paths that the MCP is allowed to access. You can include multiple entries.
• --host (optional) The host and post for the Ollama server. Defaults to http://127.0.0.1:11434
• --model (optional) The model installed into Ollama that should be used for the image vision. Defaults to "llava:34b".

Usage

To run the server:

node src/image-vision-mcp.js --permitted /path/to/dir1 /path/to/dir2

The --permitted flag is used to specify which directories roots the MCP is allowed to access for security reasons.

Development

You can run in development mode using the MCP inspector:

npx @modelcontextprotocol/inspector node src/image-vision-mcp.js --permitted /Users/FOO/Desktop/mcp/

Questions, Feature Requests, Feedback

If you have any questions, feature requests, need help, or just want to chat, join the discord.

You can also log bugs and feature requests on the issues page.

License

Project released under a MIT License.