@atercates/vision-judge

MCP server and CLI tool that analyzes web design screenshots using GitHub Copilot (GPT-4o vision) and gives direct, no-nonsense feedback on real visible problems.

Prerequisites

• GitHub Copilot CLI installed and authenticated

copilot --version  # verify installation

Installation

npm install -g @atercates/vision-judge

Tools

vision_judge_analyze

Full UI review. Use this whenever you need to evaluate how a web interface looks — and always when performing accessibility or UI checks. Returns direct feedback on visible problems, or confirms the design looks good.

vision_judge_ask

Targeted question about a specific element or aspect of the UI. Use this instead of a full review when you only need to check one concrete thing.

Examples:

• "Is the submit button visible and easy to find?"
• "Does the text have enough contrast against the background?"
• "Does the modal look centered?"

Usage

CLI

vision-judge ./screenshot.png

MCP server (Claude Desktop / Cursor / any MCP client)

{
  "mcpServers": {
    "vision-judge": {
      "command": "npx",
      "args": ["-y", "--package=@atercates/vision-judge", "vision-judge-mcp"]
    }
  }
}

Claude Code

claude mcp add vision-judge -- npx -y --package=@atercates/vision-judge vision-judge-mcp

Environment variables

| Variable | Default | Description | |---|---|---| | VISION_JUDGE_MODEL | gpt-4o | Copilot model to use for analysis |

VISION_JUDGE_MODEL=gpt-4.1 vision-judge ./screenshot.png

In MCP config:

{
  "mcpServers": {
    "vision-judge": {
      "command": "npx",
      "args": ["-y", "--package=@atercates/vision-judge", "vision-judge-mcp"],
      "env": {
        "VISION_JUDGE_MODEL": "gpt-4.1"
      }
    }
  }
}

Supported formats

PNG, JPG, JPEG, WebP, GIF — up to 20MB.