Summary

Custom models named by @ handle are the difference.

The generation call is ordinary; the part that is not is being able to name a specific LetzAI model inside the prompt, which is how you get a consistent subject or style across generations rather than a fresh interpretation each time. Upscaling accepting a bare image ID is a small convenience that adds up when you are iterating. Setup is manual — clone, compile, restart — with no npm package to install.

What it is

An MCP server for LetzAI image generation, run locally as a compiled Node script and configured in Claude Desktop with your LetzAI API key. Prompts can reference LetzAI models by their @ handle, so a generation uses a specific trained model rather than a generic one.

What you get

Image generation from a prompt, with LetzAI models referenced inline by their @ handle — the worked example is a prompt naming the `@mischstrotz` model — and the finished image opened in your browser
Upscaling of an existing LetzAI image with a strength parameter, accepting either a full image URL or just the LetzAI image ID
The generated tools visible in the client's tool list once the server is registered

Requirements

A LetzAI API key, obtained from the LetzAI API documentation and set as `LETZAI_API_KEY` in your Claude Desktop configuration. Node.js installed and on your `PATH`. The repository is built rather than installed from a registry: put the folder somewhere outside Downloads, `npm install`, then `npx tsc` to compile the TypeScript into the `build` folder, and point the config at the compiled entry point with an absolute path. Claude Desktop must be restarted after compiling and after any config change, or it will not pick up the server.

Setup effort

One command plus a key — npx tsc, then supply credentials