News

Cowork Preview Lets Claude Handle Office Tasks on Your Mac, With Risks Attached

• By John K. Waters
• 01/22/2026

Anthropic's new Cowork research preview represents a bet that the computer-using capabilities behind its developer-focused Claude Code tool can be adapted for everyday office work, giving non-technical workers a way to delegate file-heavy tasks on their Macs. The preview also highlights the risks and trust questions that arise when an AI system is granted the ability to take actions on a local machine, including the potential to delete files or be manipulated by malicious instructions embedded in content it processes.

Cowork is "a simpler way for anyone, not just developers, to work with Claude," Anthropic said in a company blog post announcing the feature. The tool is built on "the very same foundations" as Claude Code, the company’s command-line-oriented coding agent, but is designed to feel more like leaving tasks for a colleague than chatting in a standard bot interface.

Unlike a typical chatbot conversation, Cowork asks users to grant Claude access to a specific folder on their computer. From there, the assistant can "read, edit, or create files in that folder," the company said, offering examples such as reorganizing a downloads folder, turning screenshots of receipts into an expense spreadsheet, or drafting a report from scattered notes.

Anthropic said Cowork takes a more agent-like approach to work once a task is assigned. "Once you’ve set a task, Claude will make a plan and steadily complete it, while looping you in on what it’s up to," the company wrote.

The launch is the latest sign that AI companies are trying to move beyond chat interfaces toward systems that can carry out multi-step work across files, apps, and the web. Anthropic is not alone in that push: Microsoft has promoted Copilot across its productivity software for several years, and other AI labs have introduced tools for browsing the web, writing code, and interacting with software.

Anthropic is betting that a product shaped by developer usage will translate to the broader workplace. In its blog post, the company said it expected Claude Code to be used mainly for programming, but found developers "quickly began using it for almost everything else," prompting Anthropic to build Cowork for non-coding tasks.

Early coverage framed Cowork as an attempt to bring the "computer agent" concept to everyday workers, while acknowledging the persuasion problem. Microsoft’s Copilot push has not yet produced a universally agreed breakout use case, and Anthropic will need to show that delegating routine computer tasks to AI is both reliable and worth the loss of direct control.

Anthropic’s pitch is that Cowork reduces the overhead of chat-based work, where users repeatedly provide context, move files around manually, and reformat outputs. "You don’t need to keep manually providing context or converting Claude’s outputs into the right format," the company said. It added that users can "queue up tasks and let Claude work through them in parallel," rather than waiting for one request to finish.

The company also emphasized extensibility. Cowork can use Anthropic’s existing "connectors," which link Claude to external information and third-party apps. It also includes an initial set of "skills" aimed at producing documents, presentations, and other files more effectively, the company said. If paired with "Claude in Chrome," Cowork can also take on tasks that require browser access.

That broader reach is also what raises the risk profile. Anthropic cautioned that Cowork can take "potentially destructive actions," including deleting local files, if instructed, and that there is always a chance it will misinterpret a request. The company said it will ask before taking "any significant actions" and that users can limit access by choosing which folders and connectors the system can see, adding that Claude "can’t read or edit anything you don’t give it explicit access to."

Anthropic also warned about "prompt injections," a class of attacks where malicious instructions embedded in web pages, images or other content try to steer an AI agent off course. "We’ve built sophisticated defenses against prompt injections," the company said, but added that "agent safety" remains "an active area of development in the industry," recommending precautions, especially for new users.

The product is being released as a research preview, a label that signals rapid iteration and an expectation that early users will encounter rough edges. Anthropic said it plans to make "lots of improvements," including adding cross-device sync, bringing Cowork to Windows, and making the system safer.

Access is also being gated by subscription and platform. Anthropic said Cowork is available to Claude Max subscribers through its macOS app, with other users offered a waitlist. Subsequent coverage said Anthropic expanded access to include its $20-per-month Pro tier, while warning that Pro users may hit usage limits sooner than Max subscribers.

The rollout has also drawn attention because of how the tool was built. Coverage of Cowork’s development cited Anthropic’s head of Claude Code, Boris Cherny, saying the product was created in roughly a week and a half, largely using Claude Code itself. The claim has fed an online wave of jokes and unease about software that increasingly writes and operates on its own, especially among workers already anxious about automation.

Beyond memes, the product could pressure a crowded ecosystem of startups that have raised funding to solve narrow workplace problems like document drafting, file organization, data extraction, and workflow automation. Cowork’s ability to handle those tasks inside a general-purpose assistant could squeeze companies whose value is concentrated in a single feature, even as many startups argue they can still win with superior user experience or deeper workflow specialization.

Anthropic, for its part, is framing Cowork as an early step toward more practical agency, rather than a full replacement for human control. The company said users remain responsible for what they allow the tool to access, and should be explicit when tasks could cause damage, such as deleting files.

The bigger question is whether a general agent, operating through a consumer-style desktop app, can earn trust in everyday office work. Cowork is being positioned for mundane but time-consuming chores, organizing folders, converting receipts into spreadsheets, assembling drafts from notes, rather than high-stakes actions.

If it works as advertised, it could broaden the audience for agent-style AI beyond developers, who were among the earliest adopters of tools that can plan and execute tasks across codebases. If it fails, it risks becoming another demonstration that autonomy is easier to market than to operationalize safely in the messy reality of personal files, inconsistent formats and ambiguous instructions.

For now, Anthropic is asking users to treat Cowork like a powerful assistant that still needs supervision. "Stay in control," the company wrote, while encouraging experimentation to learn what the system can do and where it still falls short.