Claude Duet is Anthropic’s new iPhone agent that works after you close it

Anthropic has put Claude Duet on the App Store: a free iPhone and iPad app that triages email and calendar, runs tasks in the background, and asks before it sends, pays or signs anything. Access is limited to a waitlist.

What the app does and who can get in

The listing describes a morning routine. Claude Duet looks across email and calendar and surfaces the few items that need a decision or a reply, with context for each and a draft where it can. During the day it rechecks connected accounts every few hours and flags threads that have gone quiet, such as a client email left unanswered for days.

Users can also hand it work in a chat or by adding a task. According to the listing, it sends progress updates, keeps working after the app is closed, and returns results such as a brief, research or a draft from a single message. Work can be sorted into channels, for example one for finance and one for hiring, and Claude Duet can read across all of them. Voice is supported, either hold to talk or a spoken conversation.

It runs on the user’s existing Claude account, so memory, skills and connected apps such as Gmail and Google Calendar are present from the first message. Connected apps can be disconnected at any time.

The App Store page lists the seller as Anthropic PBC, a download size of 288.4 MB, and a Productivity category. It requires iOS 26.0 or later on iPhone, or iPadOS 26.0 or later on iPad. The only listed language is English, and the age rating is 18+. The listing shows a price of free, but downloading is not enough: the app asks for an email confirming that the account has been granted access. Signing in does not add anyone to the waitlist, which has to be joined on Anthropic’s website. Anthropic says it is adding people gradually.

Anthropic’s product page adds the conditions the App Store listing leaves out. It calls Claude Duet a research preview, says it is available on Pro and Max plans in the US, and says Team and Enterprise plans are not included yet. It also says the app shares the plan’s usage limits with the rest of Claude, so background checks and bigger tasks count against them. It lists no separate price. Besides iOS, the page mentions an Android app and a web version at claude.ai, which the App Store listing does not cover.

Approval gates and the limits Anthropic admits

The central design choice is a human checkpoint. Per the listing, Claude Duet waits for approval before it sends, posts, pays or deletes anything. Anthropic is also frank about weaknesses. In its words, “We built Claude Duet for work, so it’s weaker at personal errands.” It says the agent can make mistakes, including about what is urgent, and that something pressing can land between its periodic checks.

The listing also warns that emails and web pages written by other people can hide instructions meant to mislead the agent. Anthropic says Claude Duet is built to resist this but tells users to review what it shows them before approving. That is a notable admission for a product whose job is to read an inbox. The resistance claim is the company’s own and the listing gives no test results, so it cannot be judged from the source.

The privacy section is broad. Anthropic’s self-reported labels, which Apple notes are not verified, cover data linked to the user, including contact info, precise and coarse location, emails or text messages, photos or videos, audio data, purchase history and diagnostics. Anthropic notes that practices may vary by feature and age.

An agent that lives where users already are

The release fits a pattern of Anthropic moving Claude from a chat window into tools that act inside work systems. Last week’s Claude Tag admin API, with 14 endpoints for controlling Claude in Slack, pointed at the same shift toward agents embedded in daily workflows. Claude Duet takes that idea to the individual, on the phone.

The commercial context matters too. Anthropic has just made its IPO filing public, with a Nasdaq ticker of ANT. Tying a new agent to Pro and Max, and having it draw on the same usage limits as chat, gives subscribers a reason to stay on the higher plans and use more of their allowance. That is an inference from the product page, not something Anthropic states.

The limited rollout also looks deliberate. A waitlist, a US-only plan restriction and an English-only app let the company watch how an agent that reads personal mail behaves before widening access. The sources do not give a timeline for expansion, a price beyond the existing plans, or a date when Team and Enterprise customers might get it.

What the sources do not establish

Neither Anthropic page offers independent testing, accuracy figures or details on how often the background checks consume plan limits. The App Store page shows no ratings yet, and the app sits at chart position #23 in Productivity. How well the approval prompts and injection defenses hold up outside Anthropic’s own descriptions will only be clear once more users get in.