Feeds know exactly what you love, and that is why they keep you up. Offramp uses the same kind of knowledge to walk you down. It takes what you love from the Qloo taste graph, puts it in a calming order with NVIDIA Nemotron, and gives the night an ending.
Try an example night, or fill in your own below.
Every arrow is visible in the Agent at work panel, with the model id, the Qloo endpoint, latency and token count.
Generic sleep advice says read a book or listen to calm music. Which book, which music? Advice that is not yours is easy to ignore, and the feed is never generic. Offramp keeps the part of the feed that works (it knows you) and drops the part that hurts (it never ends, and it gets more exciting).
The server keeps nothing. The agent state travels with each request and is signed so it cannot be edited in the browser. Your morning log lives only in this browser’s storage and is sent to the model only when you press Reflect.
Every step needs an entity id from a Qloo response in the same run, and the step text is scanned for titles Qloo did not return. A draft that fails goes back to the planner with the reasons. When a Tavily key is set, one more step searches the web for a single calm episode of the first show or podcast, and keeps the pick only if its title appears in the search results.