Anthropic has launched Cowork, a new agent-style workspace that allows Claude to read, create, and edit files directly on a user’s computer. Initially released as a research preview for Claude Max subscribers on macOS, Cowork is designed to extend Claude’s capabilities beyond conversation—enabling it to complete real work across documents, spreadsheets, and folders with a higher degree of autonomy.

Cowork builds on the foundation of Claude Code, Anthropic’s developer-focused agent, but reimagines the experience for non-technical users who want AI to function more like a collaborator than a chatbot.

When Anthropic introduced Claude Code, the tool was designed primarily for software development. But users quickly repurposed it for broader tasks—organizing files, drafting documents, and managing complex workflows.

Cowork emerged from that unexpected behavior.

Unlike a traditional chat interface, Cowork allows users to grant Claude access to a specific folder on their computer. Within that scope, Claude can read files, rename them, reorganize folders, generate new documents, and synthesize information across formats.

Examples include:

This moves Claude beyond answering questions and toward executing multi-step work.

The key difference between Cowork and a traditional AI chat is not what Claude knows, but how it works once a task is assigned. In Cowork, Claude operates with greater agency than in a standard conversational session.

Once assigned a task, Claude:

This shift toward agency also changes how people interact with Claude during work. Users don’t need to wait for Claude to finish before offering additional instructions or feedback. Instead, tasks can be layered and worked through over time, making the experience feel less like a back-and-forth conversation and more like leaving instructions for a coworker.

One of the biggest friction points in using AI for real work has been the need to constantly restate context and reformat outputs. Cowork is designed to reduce that friction. Users don’t need to repeatedly provide context or manually convert Claude’s outputs into the right format.

To be useful beyond isolated tasks, an AI coworker needs access to the same information and tools people rely on every day.

Cowork integrates with Claude’s existing connectors, allowing access to external information sources where permitted. Anthropic has also added early skills that improve Claude’s ability to create structured outputs such as:

When paired with Claude in Chrome, Cowork can also complete tasks that require browser access, further expanding the scope of work it can handle.

As Claude gains the ability to act on real files and systems, the question of control becomes unavoidable. Anthropic emphasizes that Cowork operates within explicit user-defined boundaries.

Users choose:

Anthropic positions these boundaries as a core part of how Cowork operates, giving users direct control over what Claude can access. It cannot read or modify anything outside the permissions it is explicitly given, and will ask for confirmation before taking significant actions, allowing users to intervene, adjust instructions, or redirect the task.

Even with these controls in place, Anthropic notes that giving an AI system the ability to act on real files introduces new considerations. If instructed to do so, Claude can take potentially destructive actions—such as deleting local files—and there remains the possibility that instructions could be misunderstood. For that reason, Anthropic advises users to provide clear guidance, particularly when tasks involve sensitive or irreversible changes.

Users should also be aware of the risk of prompt injection, in which external content—such as information encountered on the internet—attempts to influence an AI agent’s behavior. While Anthropic says it has built defenses against these attacks, it acknowledges that agent safety, or securing AI systems that take real-world actions, remains an active area of development across the industry.

These risks are not unique to Cowork, but they may be more visible for users encountering agent-style tools for the first time. Anthropic recommends taking precautions while learning how Cowork behaves, and points users to additional guidance in its Help Center.

Anthropic is positioning Cowork as a research preview, releasing it early to understand how people actually use agent-style AI in everyday work. The company says the goal of this phase is to learn which tasks Cowork is most helpful for, where it falls short, and how it can be improved.

Anthropic encourages users to experiment with Cowork’s capabilities, including trying tasks they might not expect to work. Insights from this preview will inform future updates, which the company says are expected to include cross-device synchronization, Windows support, and additional safety improvements.

Cowork is currently available to Claude Max subscribers through the macOS app, accessible via the “Cowork” option in the sidebar. Users on other plans can join a waitlist for future access.

Q: Can Cowork take destructive actions on a user’s computer?
A: Yes. If instructed, Claude can take actions such as deleting local files within the folders it has access to. Anthropic recommends providing clear guidance and exercising caution, especially during early use.

Q: What is the risk of prompt injection in Cowork?
A: Prompt injection refers to attempts by external content—such as web pages—to alter an AI agent’s behavior. Anthropic says it has built defenses against these attacks, but notes that agent safety remains an active area of industry-wide research.

Q: Is Cowork intended for production-critical workflows?
A: Cowork is currently a research preview. Anthropic encourages experimentation but advises users to take precautions while learning how the system behaves.

Q: How does Cowork differ from traditional automation tools or scripts?
A: Unlike rule-based automation, Cowork uses Claude’s reasoning capabilities to interpret instructions, plan steps, and adapt as it works. This allows it to handle unstructured inputs—such as notes, screenshots, or mixed file types—without requiring predefined workflows.

Q: Does Cowork replace the need for human oversight?
A: No. While Cowork operates with greater autonomy than a chat-based assistant, Anthropic emphasizes that users remain responsible for oversight, especially when tasks involve sensitive data or irreversible actions.

Cowork reflects a broader evolution in how AI systems are entering everyday work. Rather than responding to prompts, tools like Cowork operate as participants in workflows—planning tasks, executing steps, and interacting directly with real files and systems.

For knowledge workers and teams, this changes the value proposition of AI. Productivity gains no longer come from better answers alone, but from delegating real work—organizing information, drafting materials, and coordinating outputs—without constant supervision. At the same time, Anthropic’s emphasis on permissions, confirmations, and safety highlights an important reality: as AI systems gain agency, human judgment becomes more—not less—important.

Cowork offers an early look at how AI collaboration may scale across organizations: not as a replacement for human workers, but as a force multiplier whose effectiveness depends on clear boundaries, oversight, and trust.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.