punyaslokdutta makes Mike run on your own hardware
A small branch swaps Mike's cloud-only AI plumbing for a setup that talks to any local model runtime - no hosted API key required.
punyaslokdutta has wired Mike's AI adapter so it can point at a local model server instead of the hosted OpenAI service. The trick: tools like Ollama and vLLM (lightweight runtimes that let you run open-source models on your own machine) already speak OpenAI's protocol, so a few configuration knobs are enough to redirect Mike at them. Crucially, Mike no longer demands a cloud API key to start up - a local runtime now counts as a real provider on its own.
The model picker also grows two Qwen options, an open-weights family from Alibaba, in mid-sized variants that fit on a serious laptop or workstation. And answers now stream token-by-token on the local path, because a blank screen while a local model thinks for a minute is the difference between usable and abandoned. A Docker-plus-Ollama walkthrough is included.
The branch is open against upstream and hasn't landed yet.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?