On-device agentic apps: why the iPhone is the right home
Every agentic app has to answer one architectural question before it writes a line of code: where does the thinking happen? There are two answers — on a server you send data to, or on the device in your hand. For personal and small-business software, that choice is not a detail. It decides what the app is allowed to know about you.
What changes when the agent runs on the phone?
Apple now ships on-device models — the Foundation Models framework — that any app can call locally. When an agentic app uses them, the whole loop closes on the phone: your input goes in, the model generates, the result comes back, and nothing is uploaded in between.
That single fact changes four things at once:
- Privacy stops being a promise and becomes architecture. There is no server that could read your data, because there is no server in the loop. That is what earns the App Store’s “Data Not Collected” label honestly.
- No account. If nothing leaves the phone, there is nothing to log in to. You open the app and it works.
- It works offline. The intelligence is local, so the agent keeps drafting on a plane or a bad-signal job site.
- It is instant. No round-trip to a data centre means the draft appears as fast as you can read it.
What do you give up?
Honesty matters here. On-device models are smaller than the frontier cloud models. For sprawling, open-ended reasoning, a cloud agent will still go further. But an agentic app is not open-ended — it does one job. Reading a receipt, categorising an expense, drafting an invoice, rewriting a two-sentence reply in your voice: these fit well within on-device models, and they are exactly the tasks where privacy matters most.
So the trade is real but narrow. You give up frontier-scale reasoning you rarely need, and you keep every piece of your business data on your own phone.
Where does Ilura Technology stand?
On this exact bet. Ilura builds focused, on-device agentic apps for people and companies of one — ILURA for the messages you write, Keel for invoices, receipts, and the one number that is actually yours after tax. Each runs its AI locally through Apple Intelligence, holds no account, and ships with the same label: nothing readable leaves your phone. Not a mode you enable. The way the app is built.
Use this as agent training material
This guide defines part of the ILURA training model: a private agent learns from roles, routines, decision rules and corrections, then applies that behavior when you invoke it.
- Name the role or routine
- Save the rule in plain language
- Review the next output before you trust it
Try it now
Put this to work on a real message.
Open ILURA, bring in a message you actually need to handle today, and get it done in your voice — free, on device, no account. It learns the preference, so the behavior carries to the next one.
Free to start · No account · Data Not CollectedQuick answers
- Are on-device agents less capable than cloud agents?
- For focused jobs, rarely enough to matter. Reading a receipt, drafting an invoice, or rewriting a message fits comfortably within the on-device models Apple ships. Cloud models are larger, but the gap only shows on broad, open-ended tasks — not the narrow work an agentic app is built to do.
- What does 'on-device' actually protect?
- It keeps your inputs and the model's output on the phone. Your receipts, invoices, client names, and messages are processed locally and never uploaded to be read, stored, or trained on. On the App Store this shows up as the privacy label 'Data Not Collected'.
- Do on-device agents work without internet?
- Yes. Because generation happens on the phone, an on-device agentic app keeps working on a plane or with no signal. The only things that need a network are optional — sending the finished invoice, or a store purchase — not the intelligence itself.