I have one that's currently still cooking, I have good experimental validation for it but I need to go back and tune the latency and improve the install story. It should help any model quite a bit but you have to hack other agents to integrate it into their api call machinery, I have a custom agent I've built that makes it easy to inject though.