Anthropic Tests ‘Cowork,’ an AI Tool That Can Act on Files and Folders

Anthropic has introduced Cowork, a new experimental feature that moves its AI assistant Claude beyond text-based conversations into hands-on task execution on a user’s computer. The tool is being released as a research preview for Claude Max subscribers using the company’s macOS app.

Unlike a standard chatbot interaction, Cowork allows users to grant Claude access to a specific folder on their device. Within those boundaries, the AI can read, edit, rename, delete, or create files. The company says early users of Claude Code, originally designed for developers, began using it for broader work tasks, prompting Anthropic to build a simpler, more accessible version aimed at non-coders. 

With Cowork, users can assign multi-step tasks such as organizing downloads, compiling expense spreadsheets from screenshots, or drafting reports from scattered notes. Once a task is set, Claude independently plans and executes the work, updating the user along the way rather than waiting for continuous prompts. Anthropic describes the experience as closer to delegating work to a colleague than holding a back-and-forth conversation.

The tool builds on the same foundations as Claude Code but is positioned for everyday productivity tasks. Cowork can also integrate with existing “connectors,” which link Claude to external sources of information, and includes new skills designed to improve document and presentation creation. When paired with Claude’s Chrome integration, the AI can also perform tasks that require browser access.

Anthropic emphasizes that user control remains central to the design. Claude can only access folders and connectors that are explicitly approved by the user and will request confirmation before taking any significant actions. 

Still, the company acknowledges risks. Since Cowork can perform potentially destructive actions, such as deleting files, users are advised to give precise instructions. Anthropic also flags ongoing industry-wide challenges around agent safety, including the risk of “prompt injection” attacks that could influence an AI’s behavior through external content.

The company says Cowork is being released early to understand how people use AI systems that operate with greater autonomy. Planned improvements include cross-device syncing, a Windows version, and additional safety measures, informed by feedback from the preview phase.

Cowork is currently available to Claude Max subscribers via the macOS app, while users on other plans can join a waitlist for future access.