This server enables non-vision AI models to analyze images and capture screenshots using OpenRouter vision models.
Analyze Images: Send an image (local file path or URL) with a text prompt to a vision-capable OpenRouter model and receive a natural-language response. Local files are validated for size and path security, then encoded as base64 before being sent.
Capture Screenshots (Windows only): Take a screenshot of a running application window by specifying either its process name (e.g. notepad) or Process ID (PID), using Windows PrintWindow via PowerShell.
Allows capturing screenshots of WezTerm windows on Windows via the screen tool, enabling analysis of terminal content through vision models.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@vision-mcp analyze this image: ./screenshot.png and explain what it shows"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server that helps non-vision models extract and understand images and screenshots via OpenRouter multimodal models.
analyze_image — Send an image (local file or URL) with a text prompt to any OpenRouter vision model and get a natural-language response.
screen — Capture a screenshot of an application window by PID or process name (Windows only, uses PrintWindow via PowerShell).
Related MCP server: Vision MCP Server
Node.js 18+ (ESM)
An OpenRouter API key with access to a vision-capable model
# Clone and install
git clone <your-repo>
cd vision-mcp
npm install
# Configure environment
cp .env.example .envEdit .env and set your OpenRouter key and preferred model:
OPENROUTER_API_KEY=sk-or-v1-...
OPENROUTER_MODEL=google/gemini-2.0-flash-lite-preview-02-05:freeVariable | Required | Default | Description |
| Yes | — | Your OpenRouter API key |
| Yes | — | Model identifier (e.g. |
| No |
| API base URL override |
| No |
| Maximum allowed image size in MB |
npm run build
npm startThe server uses stdio transport and is designed to be launched by an MCP-enabled host (e.g. Claude, OpenCode). Example configuration block for your MCP client:
{
"mcpServers": {
"vision-mcp": {
"command": "npx",
"args": ["-y", "@forloopcodes/visionmcp"],
"env": {
"OPENROUTER_API_KEY": "sk-or-v1-...",
"OPENROUTER_MODEL": "google/gemma-4-31b-it:free"
}
}
}
}analyze_imageAnalyze an image using a vision model.
Arguments:
Name | Type | Required | Description |
|
| Yes | Instruction for the vision model |
|
| Yes | Local file path or |
Local images are loaded, validated (size, path scope), and sent as base64 data URIs. Remote URLs are passed through directly.
screenCapture a screenshot of a running application window (Windows only).
Arguments:
Name | Type | Required | Description |
|
| Yes | The capture action |
|
| No | Process name to find and capture (e.g. |
|
| No | Process ID to capture |
Provide either target or pid. The server uses PrintWindow via a compiled C# snippet invoked through PowerShell. Supports PW_RENDERFULLCONTENT for DPI-aware captures.
Screen capture requires theSystem.Drawing assembly and works on Windows only. The captured PNG is written to the system temp directory.
src/
├── server.ts # MCP server entry point, tool routing
├── image.ts # Image loading, validation, base64 encoding
├── openrouter.ts # OpenRouter API client (multimodal chat completions)
├── types.ts # Shared TypeScript interfaces for API payloads
└── screen/
├── index.ts # Platform abstraction for screen capture
└── windows.ts # PowerShell + Win32 implementation (PrintWindow)The server runs on the Model Context Protocol SDK with stdio transport. Image analysis requests are forwarded to OpenRouter's /chat/completions endpoint with a 60-second timeout. Security boundaries are enforced at the image loader — only files within the working directory are accepted.
Local image paths are checked against the project root to prevent directory traversal.
Screen captures are written to the OS temp directory ($TMPDIR).
The PowerShell execution uses -ExecutionPolicy Bypass for the embedded C# script; no external .ps1 files are written to disk.
MIT
Send an image (URL or local path) with a prompt to a vision model via OpenRouter
Inputs: prompt (required), image (required).
Capture a screenshot of an application window by PID or process name (e.g. wezterm)
Inputs: action (required), target, pid.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 2 tools
The two tools have clearly distinct purposes: analyze_image sends an image to a vision model, while screen captures a screenshot. No overlap in functionality.
Both tool names follow a simple pattern, but 'screen' is a single noun used as a verb, while 'analyze_image' is verb_noun. The inconsistency is minor given only two tools.
For a vision-related server, two tools are very minimal. Typical vision servers have at least 5-10 tools for different operations (list models, describe, compare, etc.).
The server lacks essential operations like listing available models, configuring parameters, or handling multiple image inputs. The surface is too narrow for a comprehensive vision tool.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
OCR and document understanding: extract text from images, then summarize or translate it.
Image & PDF tools for AI agents: compress, convert, resize, PDF, AI vision, pipeline.
Screenshot, PDF and HTML-to-image rendering API so Claude and Cursor can see any web page.