An OpenAI-Compatible API Server on Your LAN

Turn on a small HTTP server and the model on your phone becomes available to everything else on your network, speaking the OpenAI API that half the tooling in the world already talks.

What it serves

• GET /health — a liveness probe, no authentication.
• GET /v1/models — the chat models you have installed.
• POST /v1/chat/completions — completions, streaming or not.

Point any OpenAI-compatible client at your phone's address on the LAN and it works: an editor plugin, a script, a desktop chat client, a notebook.

Why this is a genuinely odd and useful thing

It inverts the normal arrangement. Usually your laptop sends your text to a company's servers. Here your laptop sends it to the phone in your pocket, over your own network, and the phone answers. Nothing crosses your router's edge. For a script that processes something sensitive in bulk, that is a different risk profile entirely — and it costs nothing per call.

It shares one model with everything else

The phone has a single model context, so requests from the network queue on the same lock as your chat and any agents. BatteryGuard keeps its authority too: when the phone cannot afford to generate, the server answers 503 rather than pretending. An honest failure a client can retry beats a phone that overheats serving a laptop.

Scope

It is bound to your local network. This is not a way to publish a model to the internet, and it should not be made into one.