Pi works great for local models in my short testing. Wanted to try out Skills and how small models work with these agentic tools for small tasks, mainly [browser-use](https://github.com/browser-use/browser-use).
Tried Mistral Vibe, Codex, Opencode, Claude with gpt-oss:20b, ministral 3b,8b, Nemotron3 nano 30b and GLM 4.6V; finally settling on gpt for its impressive pass rates. All the other tools inject upto around 7-10k tokens on the initial prompt while pi takes up ~1.5k. This works out to be quite usable for my m3 pro machine that can take a while when processing the huge initial prompts from other CLIs.
While I'm not doing any serious work, and the other tools could be tweaked to use a simpler System Prompt; pi felt quick and the llms did use the tool calls correctly without being confused with all the huge prompts being dumped on them.