Claude: One Task, Every Screen

A browser task rarely ends where it starts. We open a tab, gather context, switch devices, and often spend the next session rebuilding the thread we just left behind.

Claude says that its Chrome sessions now carry over to desktop, web and mobile. The practical promise is continuity; the important caveat is that a page can still contain instructions designed to mislead a browser agent.

The useful change is not simply another place to open Claude. It is the chance to begin a multi-step job in Chrome and continue the same work elsewhere, with the earlier conversation still present.

That changes the rhythm of ordinary browser work. We can pick up the task rather than reconstruct it—but continuity is convenience, not a guarantee that every page in the task deserves trust.

Claude’s announcement says conversations are saved, while skills and connectors work in the browser. A second Claude post describes sessions as living with the account rather than any one device.

The rollout is narrower than a universal launch. Claude named Max and Team as available on 12 August 2026, with Pro rolling out in the following weeks; the posts do not give a completion date, price, retention policy or permission detail.

Claude’s central claim is behavioural rather than architectural: start work in a tab, then pick it up later somewhere else. That is enough to explain why the feature could make browser sessions less disposable.

It is still Claude’s own evidence. The announcement supports the continuity claim; it does not independently establish how every account, connector or browser workflow behaves in practice.

“Your Claude in Chrome sessions now carry over to desktop, web, and mobile. Conversations are saved, and your skills and connectors work in the browser.”

Claude announcement on X

The official thread provides a useful receipt for the launch: Claude has publicly described a continuous session across its surfaces. It does not provide an independent test of reliability, access conditions or connector behaviour.

The safety post matters because it keeps the claim in proportion. Claude says browser agents can be tricked by hidden instructions in a page, even as it describes defences and recommends habits of our own.

The practical sequence is simple. We begin a task in Chrome, the session remains attached to the Claude account, and we can return to that work from desktop, web or mobile.

That is a behavioural explanation of Claude’s wording, not a claim about storage, retention or underlying system design. The saved thread helps preserve context; it does not tell us what permissions a connector has or how long information is kept.

The better journey is easy to picture: begin the work in a tab, continue it from another screen, and keep the earlier context intact. That is useful when a task involves several steps rather than one isolated question.

But the moment a browser agent acts on a page, the page itself becomes part of the risk. We should inspect unexpected instructions, keep sensitive actions for human confirmation, and treat continuity as help with context—not permission to trust everything encountered online.

Claude’s update makes a browser session feel more like ongoing work and less like a temporary tab. That is a meaningful improvement when the job spans devices, conversations, skills and connectors.

The stronger reading is also the more cautious one: carrying context forward does not make web pages benign. Claude’s own safety warning leaves human checking as part of the workflow.

A session that follows us is more useful. A page that can steer an agent still needs scrutiny.

Synthesis of Claude’s continuity and browser-safety posts