The bird feeder has a research assistant

The feeder camera streams to a pipe URL all day; a daemon in the container is the receiver. Every thirty seconds it pulls one frame from the stream and hands it to a script. The script sends it to a vision model through the Hoody AI Gateway with the container's own token, asks "which species, how many," and writes a row into SQLite.

By October the table holds a season: first junco, last hummingbird, the week the jays took over. The chart is a query.

The box is flat-rate; the daemon and the database add nothing to the bill. Model calls run at token cost plus 5%, drawn from the AI Balance you topped up in April; a runaway script can't touch the server money. The Pi holds no vision key. It speaks curl.

The bird feeder has a research assistant, and it lives in a container.