Anthropic has launched Cowork, a research-preview feature that lets Claude work directly with files on a user’s computer. The new capability extends the agent architecture behind Claude Code to nontechnical tasks such as organizing folders, turning receipt images into spreadsheets, and drafting reports from scattered notes.

Cowork is available initially through Anthropic’s macOS desktop application, and only to Claude Max subscribers. The company is positioning it as a simpler way to delegate multi-step work to Claude without requiring users to operate a terminal or write code. That puts Anthropic closer to Microsoft’s productivity-focused Copilot products, while also testing whether consumers and businesses are willing to give an AI controlled access to local files.

From Claude Code to Cowork

Cowork grew out of an unexpected pattern among Claude Code users. According to Anthropic engineer Boris Cherny, developers had been using the coding tool for tasks far beyond software development, including vacation research, slide creation, email cleanup, subscription cancellations, and recovering photos from a hard drive.

Anthropic’s explanation is that the underlying agent was useful even when the task had nothing to do with programming. Cowork removes the command-line interface and presents similar capabilities through a desktop workflow intended for a broader audience.

Users select a specific folder for Claude to access. Within that designated area, the agent can read existing files, edit them, and create new ones. Anthropic has described workflows including sorting and renaming a cluttered downloads folder, extracting expense information from receipt screenshots, and assembling a first draft from notes stored across multiple documents.

The system is built on Anthropic’s Claude Agent SDK, the same underlying architecture used by Claude Code. Rather than responding with a single answer, Cowork can plan a task, perform several operations, review its work, and ask for clarification when it encounters uncertainty. Users can also queue multiple tasks, making the interaction closer to delegating work than conducting a conventional chatbot exchange.

Cowork can use Anthropic’s existing connectors to services including Asana, Notion, and PayPal when those connections are configured in Claude. It can also work with Claude in Chrome for browser-based actions such as navigating sites, filling forms, and extracting information. Anthropic has added early “skills” for tasks involving documents, presentations, and other file formats.

Evidence, claims, and an unusually fast build

The most striking development claim around Cowork concerns how quickly it was built. VentureBeat reported that Anthropic employee Felix Rieseberg said during a livestream that the team created the feature in approximately a week and a half. The same report relayed outside speculation that Claude Code may have written much of Cowork itself.

That recursive-development claim should be treated cautiously. The available evidence indicates that Anthropic employees discussed the short development timeline and that observers inferred a substantial role for Claude Code; it does not establish how much production code the model generated or how much human engineering and review were involved.

The broader product claims also come primarily from Anthropic’s own announcements and employee comments. The company says Cowork includes a built-in virtual machine for isolation, browser automation, connectors, and clarification behavior. It has described the feature as early and raw, comparable to the initial state of Claude Code. There is no independent benchmark in the available reporting showing how reliably Cowork completes file-management or office-productivity tasks, nor is there verified adoption data beyond its limited Max availability.

That distinction matters for buyers. The examples demonstrate the type of workflow Anthropic wants to enable, but they do not yet show how the system performs on messy enterprise folders, inconsistent naming conventions, sensitive documents, or tasks requiring reliable repeatability.

The access and safety trade-off

Cowork’s usefulness depends on a level of access that ordinary chat interfaces generally avoid. Once a user grants access to a folder, Claude can make changes inside it rather than merely suggesting what a person should do. Anthropic explicitly warns that the agent may take destructive actions, including deleting local files, if it is instructed to do so or misunderstands the request.

The company recommends giving clear instructions, particularly for sensitive operations. It also warns about prompt injection, in which malicious instructions hidden in a document or web page could influence the agent’s behavior. Anthropic says it has built defenses against these attacks, but acknowledges that securing agents capable of taking real-world actions remains an active industry problem.

The folder-based model is an important design choice. It limits the initial scope of access compared with an assistant that can operate across an entire operating system. At the same time, a folder can contain personal records, financial information, credentials, or proprietary business material. Isolation reduces the blast radius of a mistake, but it does not eliminate the need for backups, permissions, audit trails, and human review.

For enterprise teams, the browser and connector features create a second layer of risk. A task that begins with local documents may eventually touch external services or websites. Organizations will need to understand which credentials are used, what actions are logged, and whether approval is required before an agent changes data outside the selected folder.

What Cowork means for AI builders and enterprises

Cowork reflects a shift in the competition from model access to workflow execution. Many users can already ask an AI system to summarize a document or draft an email. The harder product challenge is allowing the system to locate the right information, manipulate files, use external tools, recover from errors, and stop when it lacks sufficient context.

For builders, Anthropic’s path from Claude Code to Cowork suggests that general-purpose agent behavior may be packaged into different interfaces rather than developed separately for every audience. A tool designed for programmers has become a foundation for office work, file management, and research. That could shorten development cycles, but it also means safety lessons from developer tools will carry into products used by less technical customers.

For enterprises, the initial Max-only rollout makes Cowork more of an evaluation signal than a deployable workplace platform. The feature is not yet broadly available to Free, Pro, Team, or Enterprise users, although people on other plans can join a waitlist. Businesses considering similar agents should test narrow workflows first: for example, generating a report from a controlled folder while preserving originals and requiring approval before deletion or external submission.

Cowork also gives Anthropic a different route into productivity software. Microsoft has pursued an operating-system and workplace-suite strategy with Copilot, while Anthropic is starting from an agent that already has developer credibility and placing it inside a desktop application. Whether that technical lineage produces better results remains unproven, but it gives Anthropic a clear product narrative: the same agent that can modify code should be able to modify ordinary work files.

What to watch next

The most important signals will be Cowork’s expansion beyond macOS and Claude Max. Anthropic has said it plans to bring the feature to Windows and add cross-device synchronization, but the timing and final scope are not yet clear.

Users and buyers should also watch for independent evaluations of file accuracy, destructive-action rates, prompt-injection resistance, and browser automation. Anthropic’s handling of approvals, activity history, undo functions, and enterprise administration will determine whether Cowork can move from a power-user preview into managed workplace deployments.

Finally, the market will be watching whether Anthropic opens Cowork to Team and Enterprise customers, and whether the company can demonstrate repeatable productivity gains rather than isolated demonstrations. The speed of the reported build is notable, but reliability and controllability will matter more than development velocity for organizations handling sensitive data.

Creati.ai perspective

Cowork is significant because it makes the agent interface tangible: users assign a job, provide a bounded workspace, and expect Claude to act across multiple files. That is a more consequential product step than adding another chat mode, but it also exposes the gap between impressive demonstrations and dependable automation.

Anthropic’s strongest advantage may be the reuse of Claude Code’s agent architecture. Its immediate challenge is proving that the same capabilities can be made safe and predictable for people who do not understand terminals, permissions, or prompt injection. Until broader access and independent performance evidence arrive, Cowork is best viewed as an important research preview—not yet a proven replacement for human-controlled productivity workflows.