stakwork teaches Mike to run on local models

The work adds support for Ollama - a tool for running open large language models on your own server rather than calling out to a hosted API - and threads it through both the backend and the chat interface. Once it's pointed at a local server, installed models show up as their own group in the model picker, with a health badge telling you whether the connection is live.

There's a deliberate trade-off baked in. Local models handle structured commands inconsistently, so when one is selected the team switches off document editing and citations and surfaces a banner explaining why. Worth knowing before you switch: the setting is server-wide today, with no per-user endpoint, so it's an all-or-nothing choice for the whole deployment.