Airtap

The arrangement is the listing. Every other agent here that acts inside apps does it through a browser or an API; Airtap gives each person a real Android phone running in the cloud, signs them into their own apps on it, and has an agent tap, type and swipe through them. You reach the agent by text, on iMessage, RCS, SMS or Telegram by its own statement, and the results land back in the thread. There is a web app to watch the phone's screen stream live and to build routines, and an Android app that can drive your own device instead, but neither is the way in.

What it is for is the chore that lives inside an app: the ready-made routines are a flight-price tracker, a LinkedIn job applier, a coupon clipper across grocery and pharmacy apps, an Instacart cart built from a meal plan, an Orangetheory class booked the moment it opens, and an Amex offers sweep. A task you liked becomes a routine that reruns on a schedule and texts you the total.

On credentials its answer is architectural: you log into your apps yourself on the cloud phone, the agent can see the screen but is, in its words, cryptographically blocked from reading password and card fields, and sessions are wiped on logout. The policy is more candid than the front page: it names six model providers that receive your task text, screenshots and extracted screen text while a task runs, says they do not train on it, and says Airtap itself may use anonymised and aggregated usage data to improve the AI, subject to consent preferences, which is a qualified answer rather than a no.

No number is published anywhere on the site. The about page says your first message creates the account, and the front page says to sign in with Google or a phone number; a reader who wants to text it first has to find the number through the web app.