Anthropic is bringing Claude Cowork to mobile and web, giving users a way to start, check, and steer longer AI-agent tasks without staying at a desktop. The beta is rolling out over the next few weeks, starting with Claude Max users before expanding to more plans.
Cowork is the Claude feature built for handing off work, not just asking questions. It can work across files, calendars, email, messaging apps, the web, and connected tools, then ask for approval when it reaches a decision only the user should make. That makes this rollout part of the same mobile-agent shift we saw when OpenAI brought Codex controls to ChatGPT on iPhone and Android.
Cowork can now keep working in the background
The biggest change is remote sessions. Anthropic says Cowork can now keep running in the cloud even after a laptop is closed, and scheduled tasks can run without any device online.
That makes the mobile app more useful than a simple status screen. A user can start a task at a desk, check progress from an iPhone, answer a question from Claude, and pick up the finished output later on another device. OpenClaw recently took a similar phone-first path with its official iPhone app for local-first AI agents.
Anthropic is also combining Claude chat and Cowork into one home on web and desktop. On mobile, Cowork is available from the sidebar in the Claude app for iOS and Android.
Web and mobile still have limits
Desktop is still the full Cowork experience. Anthropic says the desktop app remains the place where Claude can also use local files and a browser.
In simple terms, Cowork can keep a remote session alive when your computer is offline, but it cannot reach files on your Mac if the desktop app is closed. For local files, browser control, and computer-use tasks, the desktop app still acts as the bridge.
Anthropic is also extending doubled Cowork 5-hour usage limits through August 5, 2026, for eligible Pro, Max, Team, and some Enterprise users. Weekly limits are not changing.
AI agents are moving to the phone
The pattern is clear: AI agents are turning into long-running tasks that need occasional human judgment. A phone is a natural place for those approval moments, especially when the actual work is happening on a desktop, in the cloud, or across connected apps.
For Apple users, that also keeps pressure on the platform side. Apple is already reportedly looking at how the App Store should handle AI agent apps that can take actions, touch user data, or generate new behavior after installation.
Claude Cowork on mobile does not remove the need to think carefully about access. It does make the agent workflow easier to carry around. The important question now is how clearly these tools explain what they can see, what they can change, and when they will stop for approval.
