bc250-gateway Automates Power for AMD BC-250 AI Rigs
A new FastAPI gateway turns on your spare Linux box or ex-mining APU on API request and shuts it down after idle timeouts.
Running an ex-mining AMD BC-250 board or spare Linux desktop as a local AI server is cheap, but 70 watts of idle draw adds up fast. Wake-on-LAN does not work on these stripped-down boards, which have no standby power rail.
Developer atostivint released bc250-gateway, an OpenAI-compatible HTTP gateway that sits on an always-on host and controls power via a switched socket. When an API request arrives, the tool triggers a custom wake command, streams server-sent events to keep clients alive during the roughly 50-second cold boot, and proxies traffic once the backend responds to health checks.
After a configured period of inactivity, the gateway issues a graceful shutdown, monitors the host until it stops responding, and then cuts the power. It also has a single-flight wake, so concurrent requests trigger one power-on, an occupancy guard that blocks shutdowns when a user is active, and a disk-cached model list endpoint that never wakes the machine.
You can set up the FastAPI service using Python virtual environments or Docker by grabbing the code from the atostivint/bc250-gateway repository. You have to write the wake and shutdown scripts yourself.