StoweAn offline AI planner is one where the AI itself runs on your device, so planning works with no internet connection at all. Stowe is built this way: you can put an iPhone in airplane mode, speak a messy brain dump, and watch it become a sorted plan of 3–5 things — transcription, understanding, and storage all happening on the phone in your hand. Nothing you say leaves the device.
That claim is easy to make and easy to test, which is why I like it: turn the network off and use the app. This post explains what is actually happening when you do.
What happens when you speak to Stowe?
Three steps, all local. First, speech recognition: Apple's on-device speech frameworks turn your voice into text as you talk — that is the live transcript you see filling in. Second, structuring: an on-device language model (Apple Intelligence, on devices that support it) reads the transcript and splits it into items — this is a task, this is a routine, this one mentioned "Monday" so it gets a date. Third, storage: the results are saved on the phone and in your private iCloud, which syncs your own devices to each other when a connection exists and simply waits when one does not.
At no point is there a "send to server" step, because there is no server. Not "we send it encrypted", not "we delete it after processing" — there is nothing to send it to. Stowe has no accounts and no backend.
Why is a fully offline AI planner rare?
Because until recently the AI part could not run on a phone, and most existing apps were built around a server before that changed. Understanding loose human speech — "gym tomorrow, oh and mum's birthday is coming up" — takes a language model. For years, models capable of that lived in data centers, so every "AI planner" was necessarily a cloud planner, and the architecture (accounts, sync backends, server-side processing) followed from that.
On-device models like Apple Intelligence changed the constraint. A phone can now do the understanding locally. But an app built on a backend does not become offline by flipping a switch — the server is load-bearing. Building offline-first is much easier if you start there, which is what I did: Stowe was designed around the on-device pipeline from the first version, and the absence of a server is not a mode, it is the architecture.
The honest caveat: the structuring step needs a device that supports Apple Intelligence. On older devices, Stowe falls back to simpler on-device pattern matching — still private, still offline, just less clever about splitting your words.
What does on-device actually mean for privacy?
It means the privacy policy is mostly a description of what cannot happen. Your voice is processed by the phone and never uploaded. Your tasks and plans exist on your devices and in your private iCloud — the same place your photos and notes live — not in a database I operate, because I do not operate one. There is no server of mine to breach, no transcript of your life for anyone to request, and no account to identify you.
This is a different kind of promise than "we take your privacy seriously". It is structural. I could not read your brain dumps if I wanted to, and I built it that way on purpose — the app exists precisely for the thoughts you would not say into a microphone that uploads.
What does airplane mode mean practically?
It means the app's behavior does not depend on where you are. Some places this turns out to matter:
- Flights — the obvious one. A long flight is when half-formed plans surface; capture works at altitude exactly as it does at home.
- Dead zones — basements, gyms, trains, hikes, the parts of the world where coverage is a rumor. A planner that needs a signal fails exactly when you are away from your desk, which is when thoughts arrive.
- Latency — there is no round trip. Transcription is live and structuring happens on the device, so capture does not stall on a slow connection.
- Longevity of trust — an app with no server has no server to shut down. Your data's availability does not depend on my infrastructure.
The one thing a connection does add is iCloud sync between your own devices — capture on your Watch or Mac, and it is waiting on your phone. Offline, each device keeps working; they reconcile when the network returns.
Does offline mean less capable?
Less capable than a frontier cloud model, yes; less capable than the job requires, no. A data-center model can write essays. Structuring "call mum, post the passport forms, dentist Monday" into three typed, dated items does not need one — it needs a competent local model and an app that keeps the task small. Stowe also keeps every guess editable, so when the model reads something wrong, the fix is a tap, not a support ticket.
That is the design bet behind Stowe, and behind this post: for a daily planner, the ceiling that matters is not how smart the AI is, but whether you trust the app enough to actually say what is in your head. On-device is how I earn that. If you want to check any of it, the test remains the best one I can offer: airplane mode on, and talk.
Frequently asked questions
Does Stowe work without Apple Intelligence?
Yes, with a simpler brain. On devices without Apple Intelligence, Stowe falls back to on-device pattern matching for the structuring step — still private, still fully offline, just less clever about splitting loose speech into items. Transcription and storage work the same either way.
Does Stowe need an account?
No. There is no account, no sign-up, and no server to sign in to. Your data lives on your devices and in your own private iCloud, which is tied to your Apple ID, not to anything of mine.
Can a fully offline app still sync between devices?
Yes — through your private iCloud rather than a vendor server. When a connection exists, your own devices reconcile with each other; when it does not, each device keeps working alone and syncs later. Offline-first means the network is an accelerant, not a dependency.
How do I verify an app is really offline?
Airplane mode, then use every feature you care about — not just viewing your list, but capturing and processing something new. Apps that only cache reads will fail at the understanding step. Any app whose AI genuinely runs on-device will behave identically with the radios off; that test costs thirty seconds and no trust.