I do not stick to one coding agent, but pi is the one I reach for most often. It fits naturally into a terminal session and is a good default when I want to work through a change with an agent nearby. I still use Claude Code sometimes; the right choice mostly depends on the job in front of me.

herdr is part of the setup too.

There is also a local Gemma 4 cluster on the network. LiteLLM sits in front of it, so clients can use it through one endpoint rather than each needing to know about the cluster directly. It is useful to have a local model available without making it a separate workflow.