Claude Cowork Comes to Mobile and Web for Background Tasks

Anthropic built its reputation on Claude Code, a terminal agent that let developers hand off real programming work and walk away. The company has spent 2026 arguing that the same idea belongs far outside engineering, and its latest release makes the point hard to ignore. Claude Cowork, the general-purpose version of that agent, no longer lives only on a Mac desktop. As of July 7, 2026, it runs on mobile and web too, which means the assistant can keep grinding through a task after you shut the laptop and tap you on the shoulder, through your phone, when it hits a decision only a human should make.

The rollout was reported by TechCrunch and The Decoder, both of which framed it as a step change in how the agent fits into a working day. Until this week, Cowork was tethered to the machine it ran on. Now the task and the person minding it can sit on different devices, in different places, which is a bigger shift than the plain feature list suggests.

From desktop preview to your pocket

Cowork started life narrowly. Anthropic introduced it in January 2026 as a research preview inside the Claude macOS desktop app, pitched as a computer agent for people who are not software engineers. As NBC News noted at the time, the appeal was permission-based access to local folders: rather than uploading files into a chat window, users could point Cowork at a directory and let it read, edit, and create documents directly on the machine. It was Claude Code's engine repointed at spreadsheets and slide decks instead of source trees.

That desktop tether was also the ceiling. An agent that only exists inside one app on one laptop cannot run an errand while you commute, and it cannot let you glance at progress from a meeting room. The July update lifts that ceiling. According to The Decoder, beta access goes live gradually over the coming weeks and starts with Max subscribers, the top consumer tier. TechCrunch confirmed the same Tuesday launch and the same Max-first availability, with SiliconANGLE describing it plainly as bringing Cowork off the desktop and onto web and mobile.

Anthropic also sweetened the launch for early users. The Decoder reports that usage limits were doubled through August 5, a common tactic to get people leaning on a new surface hard enough to form a habit before the caps tighten again. For a feature whose whole value depends on being trusted with real, ongoing work, generous early limits are less a promotion than a bet that once people see the background execution work, they will not want to go back to babysitting a chat window.

Gating the launch behind the Max tier is its own signal. Max is Anthropic's most expensive consumer plan, and putting the most autonomous version of Cowork there first tells you who the company thinks will pay for an agent that runs on its own. These are heavy users, the people whose workflows already justify a premium subscription, and they are the group most likely to hand a tool genuinely valuable work rather than test prompts. Starting there also gives Anthropic a controlled pool to watch as the cloud version scales, since a smaller, more committed audience surfaces problems before a wider release would. If the background model holds up with Max subscribers leaning on it hard, a broader rollout down the tiers becomes a much easier decision to justify later.

How Claude Cowork works across devices

The mechanic that makes the expansion matter is continuity. You can kick off a task at your desk, check on it from your phone, and open the finished result in a browser somewhere else entirely. The Decoder describes the agent as continuing to work in the background even when the laptop is closed or the phone is switched off, because the execution no longer depends on any single device staying awake. The work happens in Anthropic's cloud, and your devices are just windows onto it.

The more interesting design choice is how the agent handles its own limits. Cowork does not simply barrel ahead and hope. When it reaches a point where only a person can make the call, it pauses and asks, and it does so through a smartphone notification rather than waiting silently for you to return to the app. The Decoder captured the pattern cleanly: when Claude hits a moment that needs human judgment, the agent asks via the phone. That turns a long-running task into something closer to delegating to a junior colleague who texts you a quick question and then keeps going once you answer.

This is a meaningful change to the rhythm of using an AI tool. Chat assistants reward constant attention, because the model waits for your next message and the session stalls the moment you look away. An agent that reaches out to you inverts that relationship. You set the objective, go do something else, and the tool reappears only when it has earned your attention. TechCrunch summarized the everyday version of this as starting a task from your desk, getting status updates on your phone, and picking up the output later even with the laptop closed.

Most people using Cowork are not coding

The most revealing part of the announcement is not the plumbing but the usage data, which quietly upends the assumption that agents are mainly a developer toy. Anthropic says more than 90 percent of what Cowork does is not software development at all. That figure alone reframes the product. The Claude Code lineage suggests a coding tool, but the people actually using its general-purpose sibling are mostly doing office work.

VentureBeat and TechCrunch both cited a study Anthropic ran on 1.2 million anonymized sessions from more than 600,000 organizations during May 2026. The breakdown is stark. Business process operations accounted for 33.4 percent of usage, content creation and copywriting for 16.4 percent, and software development for just 8.7 percent. Put differently, the single largest use of a tool descended from a coding agent is administrative operations work, and coding is a distant third.

The concrete examples make the categories feel real. The Decoder lists tasks like reconciling quarterly spend, drafting variance memos, and building client decks from call transcripts. TechCrunch describes people pulling scattered updates into a single report, building onboarding checklists, then reconciling spreadsheets that never quite agree. These are the unglamorous jobs that eat hours of a knowledge worker's week, the kind nobody puts on a resume but everybody has to do. An agent that can be handed the messy middle of that work, then checked from a phone, is aimed squarely at those hours.

Anthropic has leaned into this positioning. Rather than calling Cowork a coding assistant, the company frames it as an agentic administrative coworker, a system meant to sit where the actual work happens instead of behind a chat prompt. The name itself does the work: Cowork, not Code. The bet is that the total addressable market for an agent that drafts memos and reconciles spreadsheets is far larger than the market for one that only writes software.

What the web version cannot do yet

The expansion comes with real tradeoffs, and Anthropic has not buried them. The Decoder reports that the web version lacks several of the desktop app's most powerful capabilities. It cannot read and write local files directly. It drops plugin support, and it gives up browser control along with Computer Use, the feature that lets Claude operate a graphical interface by moving a cursor and clicking like a person would.

Those omissions matter because they are exactly the abilities that made the desktop preview feel like a genuine computer agent rather than a smarter chatbot. Local file access was the original selling point, the thing that let Cowork touch the documents already on your machine without a manual upload dance. Computer Use is what lets an agent operate software that has no clean API. Stripping both from the web build means the mobile and browser experiences are, for now, a lighter version of the tool, better suited to work that lives in the cloud than to work anchored to a specific desktop.

That gap is likely temporary, and it reflects a hard technical reality rather than a lack of ambition. Reaching into a user's local file system or driving their screen from a browser tab or a phone raises security and permission problems that a native desktop app sidesteps by running on the machine itself. Anthropic appears to be shipping the parts that translate cleanly to the cloud first, and holding back the device-bound powers until it can offer them safely. The result is a split personality worth understanding before you rely on it: the desktop app remains the fullest version of Cowork, while web and mobile trade some reach for the freedom to leave your laptop behind.

The trust problem an office agent has to solve

An agent that touches business documents inherits a harder problem than one that only writes code in a sandbox. A variance memo can contain confidential numbers. A client deck can hold information covered by a contract. When an agent reconciles quarterly spend, it is reading exactly the kind of data a company most wants to control. That is why the permission model behind Cowork is not a footnote but the load-bearing wall of the whole product, and it helps explain the careful, tiered way Anthropic is releasing the cloud version.

The original desktop preview handled this by asking for explicit, folder-level permission before it touched anything, keeping the work on the user's own machine where the security perimeter was already understood. Moving execution into the cloud changes the calculus. Now the documents an agent works on may pass through Anthropic's infrastructure, and the questions a security-conscious buyer will ask multiply. Where does the data live while a task runs. Who can see it. What happens to it when the task ends. Anthropic has not published a full answer alongside this launch, and the enterprise appetite for cloud agents will hinge partly on how convincingly it does.

The human-in-the-loop design is part of the answer. By pausing for a decision rather than pressing ahead, Cowork keeps a person on the hook for the consequential choices, which is as much a governance feature as a usability one. An agent that never asks is an agent that can quietly do the wrong thing at scale. One that stops to check gives a worker, and the company behind them, a natural point to catch a mistake before it ships. That restraint may look like friction, but for the operations and finance work that dominates Cowork's usage, it is closer to a requirement than a nicety.

The coding-agent wars spill into the office

TechCrunch put the broader story in its headline, casting this as the coding-agent wars spilling into the rest of the office, and the framing fits. The playbook is now visible across the industry: build an agent that developers trust with real work, prove the harness holds up under load, then repoint it at the far larger world of general knowledge work. Anthropic is running that play with Cowork, and it is not alone.

OpenAI has moved the same direction with Codex, which began as a software development tool and has been pushed toward broader business tasks like report generation and data analysis. The two companies are converging on the same target from the same starting point, each betting that the muscle they built for coding agents transfers to the office. Anthropic has also been widening its footprint in other directions, having recently launched Claude Tag, an always-on teammate that lives inside Slack, so that the agent can act where conversations already happen rather than waiting inside a separate app.

Seen together, these moves sketch a clear direction of travel for AI tools in 2026. The chatbot, the interface that defined the first wave of generative AI, is starting to look like a transitional form. What comes next is the persistent agent: a system you delegate to rather than converse with, that runs while you are away and pings you when it needs a decision. Anthropic's own product line now mirrors that shift, with Chat for the moments you want a quick answer and Cowork for the tasks you would rather hand off entirely. The mobile and web launch is what makes the second half of that pairing usable in the flow of a normal day. For readers following how these agentic tools are reshaping everyday software, it is one of the clearer signals yet of where the category is heading, a thread we keep tracking across the MaShop blog.

A quiet redefinition of the assistant

It would be easy to read this update as a routine platform expansion, one more feature reaching one more screen. That undersells it. Moving Cowork to mobile and web changes what the tool is for, not just where it runs. A desktop-bound agent is something you supervise. A cloud agent that texts your phone is something you employ. The difference is the gap between a smarter piece of software and a delegate you actually trust with a job.

The usage numbers are the part worth sitting with. When more than 90 percent of a coding agent's descendants are being used for operations and content work, the market is telling the labs something about where the real demand sits. The value is not in a marginally better chat reply. It is in taking a whole tedious task off a person's plate and returning a finished result, with only a quick question or two along the way. Anthropic has now made that possible without asking anyone to stay chained to a desk, and the caps it doubled through August 5 suggest it very much wants people to find out how that feels. Whether the lighter web build can carry the full weight of that promise, or whether the desktop app remains the place where Cowork truly earns its keep, is the question the coming weeks of this beta will answer, and it is the one worth watching most closely as the rollout widens.