Anthropic has taken a significant step toward making artificial intelligence a day-to-day work companion with the release of Cowork, a new tool that extends the capabilities of Claude into semi-autonomous knowledge work.
Launched as a research preview for subscribers to the company’s top-tier plan, Cowork reflects how users have already been stretching Claude Code well past its original remit. After the company released Claude Code in 2024, developers quickly began using it for tasks that had little to do with programming: organizing files, researching trips, cleaning inboxes, or assembling documents. Rather than steering those users back toward code, Anthropic decided to meet them where they were.
A Helpful Colleague
Cowork is built on the same agentic foundations as Claude Code, but strips away the command-line interfaces and developer-centric setup. Instead, users grant Claude access to a specific folder on their computer. Within that sandbox, the model can read, edit, create, rename, or reorganize files, carrying out multi-step tasks with a degree of autonomy that goes well beyond a standard chat interaction.
In practice, that means Cowork can turn a pile of receipt screenshots into an expense spreadsheet, reorganize a cluttered downloads folder, or draft a report from scattered notes, all while explaining what it plans to do and checking in before major actions. Tasks can be queued up in parallel, reducing the sense of constant back-and-forth that often defines AI-assisted work. As Anthropic describes it, interacting with Cowork feels less like chatting with a bot and more like leaving instructions for a colleague.
Under the hood, Cowork draws on the same Claude models and agent SDK as its coding-focused sibling, but adds new skills aimed at document creation and presentations. It can also tap into existing connectors that link Claude to third-party services, and when paired with the Chrome browser extension, it can complete tasks that require web access.
Notably, Cowork was built in under two weeks, using (of course) Claude Code. In a post on X, Boris Cherny, Anthropic’s head of Claude Code, touted the speed of development as proof of how broadly the agentic approach could be applied once the tooling was in place.
That power, however, comes with risks. Because Cowork can take action on local files, Anthropic is explicit about the need for caution. The system will only operate within folders users approve, and it asks before performing significant actions. Still, it can carry out potentially destructive tasks, such as deleting files, if instructed to do so. A careless user or ambiguous prompts raise the risk of serious mistakes.
Anthropic also highlights the ongoing industry challenge of prompt injection, where malicious content embedded in files or web pages attempts to manipulate an AI agent’s behavior. While the company says it has built defenses against such attacks, it acknowledges that securing real-world AI actions remains an active area of research.
Reaching A Larger User Base
Cowork arrives amid a major push across the industry to make AI agents more useful to non-technical users. Microsoft, for example, has spent years promoting Copilot as an everyday assistant, with mixed adoption. For Anthropic, apparently the hope is that the strong reception to Claude Code among developers will translate into trust among a wider audience.
For now, access is limited. Cowork is available only to Claude Max subscribers through Anthropic’s macOS app, with a waitlist for other users. The company says the limited release is intentional, allowing it to observe how people actually use the tool before expanding to Windows and adding features like cross-device sync.
If the preview succeeds, Cowork could signal a shift in how AI tools are positioned: not as chatbots waiting for prompts, but as semi-autonomous coworkers embedded directly into everyday workflows.
