manueljpconde plugs Mike into self-hosted local models
A fourth model option lets self-hosters point Mike at a model running on their own hardware instead of a cloud provider.
The fork adds a 'local' provider that talks to any OpenAI-compatible endpoint - meaning tools like Ollama or LM Studio, which let you run open-weight models on your own server. The switch is backend-only: the admin sets the model and credentials in environment variables, and the browser only ever learns a friendly label and whether it's turned on. Secrets stay on the server.
The interesting work is in the guardrails. Many self-hosted models can't reliably use tools, so when tools are off, the system prompt explicitly warns the model not to pretend it did. The tabular-review feature, which depends on tools, is hard-blocked from running on a local model unless the admin confirms tool support - quietly falling back to a known-good model instead. The dropdowns hide the local option entirely until the backend says it's ready.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?