Summary
One tool, one job — and three places the model can come from.
Precedence is explicit: the model argument on the call wins, then OPENROUTER_MODEL, then anthropic/claude-3.5-sonnet. Images are resized and converted to JPEG on the way out, so you are not shipping a full-resolution file to answer a question about a screenshot.
What it is
A single-purpose MCP server that sends a local image to an OpenRouter vision model and returns the answer.
What you get
- One tool, analyze_image, taking image_path and an optional question
- Per-call model override — pass model to use a different vision model for one image
- Automatic resizing, JPEG conversion and quality optimization before upload
- Descriptive errors for bad paths, missing keys, network failures and invalid models
- Anthropic/claude-3.5-sonnet and anthropic/claude-3-opus-20240229 are the tested models
Requirements
An OpenRouter API key in OPENROUTER_API_KEY, with OPENROUTER_MODEL as an optional default. The README's install line names @catalystneuro/mcp_read_images, which npm does not have; the command your client runs is read_images.