I have been frustrated with how repetitive the output of LLMs is. Even at maximum temperature, asking the same prompt ten times gives functionally the same result each time, as discussed here: https://news.ycombinator.com/item?id=50090443

That’s not surprising considering how LLMs work, but the thread above got me wondering if you could push models away from their usual generation patterns by injecting entropy from an external source.

After a lot of tweaking and experimenting, I turned the idea into an MCP server. It's optimized to add just enough entropy that creative output stops being repetitive, while still living up to the intent of the original prompt.

I’d love feedback, particularly on the approach itself. If you have a better way to inject entropy than what I have found, a pull request is very welcome: https://github.com/modelsalt/model-salt