Claude Code computer use turns Dispatch into a remote operator

Anthropic did not just add another clever demo to Claude this week. It connected three surfaces that used to feel separate: the phone in your pocket, the persistent thread that follows you around, and the desktop session with access to your real tools. In Anthropic’s March 23 launch post, the company said Dispatch now works with computer use so you can assign a task from your phone and let Claude operate your computer while you are away.

That sounds like a feature update. It is closer to a control-plane update.

The interesting step is not that Claude can click buttons. Plenty of AI companies have shown GUI automation demos. The interesting step is that Anthropic is now tying phone-originated instructions to a desktop session that can touch your local files, your configured connectors, your dev tools, and, when needed, the apps on your actual machine. That makes Claude Code and Cowork feel less like assistants waiting in a chat box and more like permission-mediated remote operators.

If that sounds slightly unsettling, good. Anthropic’s own docs read like they know exactly what kind of product they just shipped.

What Anthropic actually launched

The official support material for Dispatch describes it as one continuous conversation with Claude that you can reach from your phone or your desktop. Instead of opening a fresh session for each task, you keep a persistent thread. When you ask Claude to do something, it decides what kind of work is needed and spins up the right session underneath: knowledge work in Cowork, development work in Claude Code. The result comes back into the shared thread, with a push notification when the job is done or when Claude needs your approval.

That setup already mattered before computer use arrived, because it turned Claude into something closer to a delegated work queue. The new piece is what Anthropic added on top. In the same launch post, the company says Claude can now use your computer in Cowork and Claude Code with “no setup required,” opening files, using the browser, and running dev tools automatically when a more direct integration is not available.

The Claude Code desktop docs make the product shape even clearer. Dispatch integration means you can send a task from your phone and get a Claude Code session on your desktop. If computer use is enabled, those Dispatch-spawned Code sessions can use it too. In other words: your phone can now trigger work on your real development machine.

That is the important mental model. This is not just mobile chat plus a separate desktop assistant. It is a cross-device chain of command.

Why Dispatch changes the meaning of computer use

Computer use on its own is still just a capability. Useful, sometimes awkward, often slower than the clean tool path. Anthropic more or less says this openly. The docs describe computer use as the broadest and slowest option, and they spell out the priority order: Claude uses a connector first if one exists, then Bash for shell tasks, then browser tooling where available, and only then falls back to computer use.

That detail matters because it tells you what Anthropic thinks it is building. The product is not “Claude, but now with screenshots.” The product is one operator surface over several action layers: connectors, terminal tools, browser automation, and finally desktop GUI control when the cleaner route runs out.

That is why Dispatch is such a consequential pairing. It gives that action stack a remote trigger. You can be on the train, send a task, and let Claude decide whether the job belongs in Cowork, Claude Code, a connector, the shell, the browser, or the desktop itself. It is the same broader shift we wrote about in AI’s new battlefront is action, not answers, just in a much more personal form factor. Anthropic is not only chasing more capable answers. It is chasing a workflow where your intent routes into real actions on your own machine.

That is a bigger category move than the launch-video version of this story suggests.

The trust boundary is the story

The best line in the support docs is also the bluntest. Anthropic warns that “giving a mobile AI agent remote control of a desktop AI agent creates a chain” where instructions from your phone can trigger real actions on your computer, including reading, moving, or deleting local files, interacting with connected services, controlling your browser, and using desktop apps.

That is unusually plain language for an AI launch, and it is welcome.

Anthropic is also explicit that computer use in Cowork runs outside the virtual machine Cowork normally uses for file work and command execution. The model takes screenshots to understand what is on screen. It asks for permission before accessing each application. Some apps are off-limits by default. The company recommends starting with apps you trust and avoiding sensitive data. In the desktop docs, Anthropic adds another important detail: in Dispatch-spawned Claude Code sessions, app approvals expire after 30 minutes and re-prompt, rather than lasting for the whole session.

That does not make the system safe in some magical sense. It does make the real product boundary legible. You are not granting an abstract model some vague future capability. You are authorizing a chain of surfaces, sessions, and tools that eventually lands on the machine sitting on your desk.

And that is exactly why this launch has more strategic weight than a typical “AI agent can click around” announcement. The hard part is not the click. The hard part is permission, routing, and containment. Or, to put it less politely, this is where you learn whether the grown-ups were in the room when the workflow was designed.

Why this matters beyond Anthropic

Anthropic is not alone in pushing toward agents that do things instead of merely describing what you should do next. But its version is notable because it is anchored to the user’s existing machine rather than only to a hosted cloud environment. That makes the product more immediately useful, and more invasive, at the same time.

OpenAI’s agent stack, as we argued in OpenAI’s agent stack is a distribution play, not a demo, is partly about owning the default surface through which other tools are called. Anthropic is now making a similar bet from another direction: if Claude can sit across phone, desktop, connectors, terminal, and GUI control, it becomes the interface through which real work gets delegated.

The security and governance implications follow from that architecture. Once the agent can cross those layers, the crucial question becomes who owns the approval boundary and how visible the control plane stays to the user. That is why Anthropic’s launch also rhymes with the argument in NVIDIA OpenShell makes agent security an infrastructure layer. The market is moving toward agents that can act. The next competitive layer is not only model quality. It is whether action is mediated by permissions, policy, and sane runtime boundaries instead of vibes.

Anthropic’s docs, to their credit, keep pointing back to that reality. Computer use is still a research preview. It is currently available for Pro and Max subscribers. The computer-use layer is macOS-only for now, and the desktop app has to be running and awake. In other words, this is not yet a universal remote for every machine in your life. It is an early version of something narrower and more concrete: a persistent, cross-device agent that can operate your actual desktop when you allow it.

What to watch next

The next question is whether users decide this is a convenience feature or a new default workflow. If people really do start assigning morning briefings, spreadsheet edits, bug fixes, tests, and pull-request chores from the phone and retrieving results from Claude Code later, the category changes. Claude Code stops being just a coding assistant with a nicer shell. It becomes a remote operator surface.

That is the real story here. Dispatch plus computer use do not just make Claude more capable. They make Claude Code much harder to describe as “just” a coding tool.