How AGI, Inc. Keeps Its Action AI Private by Design

Updated

AGI, Inc.'s privacy model rests on one architectural choice: AGI-0, its action model, lives on the device rather than in the cloud. Because the agent sees the screen and acts across apps directly on the phone, it doesn't need APIs or third-party integrations to reach your data — and the company frames the result as "private by design." This matters most if you want an agent that can operate across sensitive apps without shipping your screen contents to a remote service.

What "on-device" actually changes

The company describes AGI-0 as a "small action model" that "lives on the device." That single fact drives the rest of the privacy story:

  • No API layer to expose. The agent works by seeing the screen and acting across any app on Android, rather than by connecting to services through APIs or integrations. Fewer integration endpoints means fewer places where credentials or data are handed to a third party.
  • Screen content stays local. Since the model runs where the screen is, the observation-and-action loop happens on the phone itself. The site's own framing is that data stays on the user's device.
  • It works offline. AGI, Inc. states AGI-0 "works offline, on wearables, and privacy-sensitive tasks" — an offline capability is only possible if the core action loop doesn't depend on a remote server.

A concrete example

Take the site's own demo prompt: "Find the playlist Jake sent me in Messages and play it on Spotify." A cloud-integrated agent would typically need read access to your messaging app and a Spotify connection. AGI-0 instead reads the screen, finds the message, and drives the Spotify app the way you would. The message content and your account aren't passed through an integration — the action happens on the device.

Where privacy and capability meet

The design isn't only about withholding data; it's also what enables certain tasks. The company positions AGI-0 for "privacy-sensitive tasks" and for wearables, where a cloud round-trip would be impractical. It also says the model "learns your preferences and gets better over time" while living on the device — so personalization is described as happening locally rather than by training on a server-side profile.

Claim on the site What it implies for privacy
AGI-0 lives on the device The action loop runs locally, not in the cloud
No APIs or integrations No third-party connection layer holding your credentials
Works offline Core actions don't require sending screen data out
Suited to privacy-sensitive tasks The company targets exactly the use cases cloud agents handle poorly

What the page does not establish

The site is a product and marketing page, so treat the privacy claims as the company's stated design intent, not an audited guarantee. It does not publish a technical whitepaper on the page, detail what telemetry (if any) leaves the device, or describe encryption or data-retention practices. If your decision depends on those specifics — for example, deploying it in a regulated environment — ask AGI, Inc. directly through the "Book a call" or "Get access" routes on the site before relying on the on-device claim.

Who this design suits

Choose this approach if your priority is keeping screen-level activity on the phone and you want an agent that reaches apps without you wiring up integrations. Look elsewhere — or get answers in writing first — if you need documented compliance controls, third-party security audits, or clarity on any data that may still leave the device. The privacy-by-design pitch is coherent and specific about mechanism; the open questions are about verification, not architecture.