An AI bedtime app with no server
July 2026 · Markus Skov
Fable writes my daughter a new bedtime story every night. Her name is in it, her stuffed fox is in it, and it always ends with her falling asleep. The part that surprises people is not the AI. It is that the app has no server.
Not "we minimize data collection". No server. There is no account system, no API endpoint, no analytics SDK, no crash reporter phoning home. The App Store privacy label is one line: Data Not Collected. If you reinstall the app, your library is gone, because we could not read it if we wanted to. That is a real trade-off, and we tell families about it in the app instead of hiding it.
How does a generative AI app work with no backend? Apple Intelligence ships a language model inside the operating system. Fable hands it a carefully fenced prompt (the child's name and the parent's words are data, never instructions), receives a structured story back, and then treats that story as untrusted input: every page passes a language-aware content gate before a child ever sees it. Scary words, wrong-language endings, missing wind-downs, stories that do not land in sleep: all rejected.
Rejected into what? That is the second architectural decision. Underneath the model sits a shelf of hand-written, parameterized stories in seven languages, deterministic under a seeded generator, edited to publishable quality. If the model is unavailable, slow, or writes something the gate refuses, the shelf answers instead, silently. Beneath the shelf sits an emergency story built only from neutralized values, and beneath that a single constant story that cannot fail. Four floors, and a child only ever sees the top one that worked. There are no error states at bedtime, because at 8pm with a tired three-year-old, an error dialog is a product bug no matter how pretty it is.
Covers work the same way: Image Playground paints them on the device from prompts the app fully controls, and if it cannot, the story keeps its quiet emoji emblem. Narration is on-device speech synthesis, and with Personal Voice a parent can lend the app their own voice, which also never leaves the phone.
The result is an AI product whose entire threat model fits in a sentence: nothing goes in but the OS, and nothing comes out at all.
If you want the safety half of this story, it has its own post: The safety gate: shipping model output to a three-year-old.