You need a local AI model because ollama doesn't work without modification, as gnosh.ai changes the way the model is used.
In the link https://gnosh.ai/head-to-head.html you can see roughly how it works. This makes all existing harnesses lossy inefficient tools in comparison.
Also see https://gnosh.ai/peers.html for multi-user parallelism.
@jaekwon and @gnosh_ai on X.