The launch marks a major inflection point in the race to deliver practical AI agents to everyday users, positioning Anthropic to compete not only with OpenAI and Google in conversational AI, but also with Microsoft’s Copilot in the burgeoning market for AI-enabled productivity tools.
“Cowork lets you complete non-technical tasks just like how developers use Claude Code,” the company announced via its official Claude account on
Over the past year, the industry has focused on large language models that can write poetry or debug code. With Cowork, Anthropic is betting that the real enterprise value lies in AI that can open a folder, read a messy stack of receipts, and create an organized expense report without the need for human control.
How developers using the holiday research programming tool inspired Anthropic’s latest product
The genesis of Cowork lies in Anthropic’s recent success with the developer community. In late 2024, the company released Claude Code, a peripheral tool that allows software engineers to automate routine programming tasks. The tool was a huge success, but Anthropic noticed a strange trend: users were forcing the programming tool to perform non-programming work.
According to Boris Cherny, an engineer at Anthropic, the company noticed that users were deploying the developer tool for an unexpected variety of tasks.
Cherny wrote on
Recognizing this use of shadows, Anthropic has effectively stripped the complexity of the command line from its developer tool to create an easy-to-use interface. In its blog post announcing the feature, Anthropic explained that developers “quickly started using it for almost everything else,” which “led us to create Cowork: a simpler way for anyone – not just developers – to work with Claude in the same way.”
Within the folder-based structure that allows Claude to read, edit, and create files on your computer
Unlike a standard chat interface where the user pastes text for analysis, Cowork requires a different level of trust and access. Users designate a specific folder on their local machine that Claude can access. Within this sandbox, the AI agent can read existing files, modify existing files, or create entirely new files.
Anthropic offers several illustrative examples: reorganizing a cluttered Downloads folder by intelligently sorting and renaming each file, creating a spreadsheet of expenses from a collection of receipt screenshots, or drafting a report from notes scattered across multiple documents.
“In Cowork, you give Claude access to a folder on your computer. Claude can then read, edit, or create files in that folder,” the company explained on X. “Try it to create a spreadsheet from a stack of screenshots, or produce a first draft from scattered notes.”
The architecture is based on what is known as the “proxy loop”. When a user assigns a task, the AI doesn’t just generate a text response. Instead, you formulate a plan, execute the steps in parallel, check their work, and ask for clarification if you hit a roadblock. Users can queue multiple tasks and let Claude process them simultaneously – Anthropic describes the workflow as “a lot like leaving messages back and forth and a lot like leaving messages for a coworker.”
The system is built on Anthropic’s Claude Agent SDK, which means it shares the same basic architecture as Claude Code. Anthropic notes that Cowork “can do many of the same tasks that Claude Code can, but more easily for non-code tasks.”
Recursive loop in which AI builds AI: Much of Claude Cowork’s work is said to have been written by Claude Cowork
Perhaps the most notable detail surrounding Cowork’s launch is the speed with which the tool was created, highlighting the iterative feedback loop as AI tools are used to build better AI tools.
During a live stream hosted by Dan Schipper, Anthropic employee Felix Risberg confirmed that the team built Cowork in about a week and a half.
Alex Volkoff, who covers AI developments, expressed astonishment at the timeline: “Holly Anthropic built Cowork in the last… week and a half?!”
This sparked immediate speculation about how much work Claude Code had built himself. Simon Smith, executive vice president of generative AI at Klick Health, put it bluntly on
The implication is deep: Anthropic’s AI crypto agent may have contributed significantly to the building of a non-technical sister product of its own. If true, it’s one of the most clear examples yet of using AI systems to accelerate their development and scale – a strategy that could widen the gap between AI labs that successfully deploy their customers internally and those that don’t.
Connectors, browser automation, and skills extend Cowork beyond the local file system
Coworking does not work in isolation. This feature integrates with Anthropic’s existing ecosystem of connectors – tools that connect Claude to external information sources and services like Asana, Notion, PayPal, and other supported partners. Users who have configured these connections in the standard Claude interface can leverage them in Cowork sessions.
Additionally, Cowork can pair with Claude in Chrome, Anthropic’s browser extension, to perform tasks that require web access. This combination allows the agent to navigate through websites, click buttons, fill out forms, and extract information from the Internet – all while running from a desktop application.
“Cowork includes a number of new user experience and safety features that we think make the product really stand out,” Cherny explained, highlighting the “built-in virtual machine.” [virtual machine] For isolation, out-of-the-box support for browser automation, support for all claude.ai data connectors, and asks for clarification when unsure.
Anthropic also introduced an initial set of “skills” designed specifically for Cowork that improve Claude’s ability to create documents, presentations, and other files. These are based on the Skills for Claude framework that the company announced in October, which provides specialized instruction sets that Claude can load for certain types of tasks.
Why does Anthropic warn users that its AI agent might delete their files?
Going from a chatbot that suggests edits to an agent that makes edits comes with significant risks. In theory, an AI that can organize files could delete them.
In a notable show of transparency, Anthropic devoted significant space in its ad to warning users of the potential risks of Cowork – an unusual approach for a product launch.
The company explicitly admits that Cloud “can take destructive actions (such as deleting local files) if asked to do so.” Because Claude may sometimes misinterpret instructions, Anthropic urges users to provide “very clear instructions” about sensitive operations.
Even more troubling is the risk of injection attacks – a technique in which malicious actors embed hidden instructions into content a cloud might encounter online, potentially causing the agent to bypass safeguards or take malicious actions.
“We have built sophisticated defenses against instantaneous injection, but customer safety-that is, the task of securing Claude’s actions in the real world-remains an active area of development in the industry,” Anthropic wrote.
The company described these risks as inherent to the current state of AI agent technology and not unique to Cowork. “These risks are not new with Cowork, but this may be the first time you’ve used a more advanced tool that goes beyond a simple conversation,” the announcement notes.
Anthropic’s desktop proxy strategy poses a direct challenge to Microsoft Copilot
The launch of Cowork puts Anthropic in direct competition with Microsoft, which has spent years trying to integrate Copilot AI into the fabric of the Windows operating system with mixed adoption results.
However, the anthropic approach differs in its isolation. By restricting the agent to specific folders and requiring explicit connectors, they attempt to balance the utility of the OS-level agent with the security of a sandbox application.
What distinguishes the anthropic approach is its bottom-up development. Instead of designing an AI assistant and tweaking the agent’s capabilities, Anthropic built a powerful coding agent first – Claude Code – and is now abstracting its capabilities for broader audiences. This artistic pedigree may give Cowork a more robust agent demeanor from the start.
Claude Code has generated significant enthusiasm among developers since its initial launch as a command-line tool in late 2024. The company expanded access with a web interface in October 2025, followed by a Slack integration in December. Cowork is the next logical step: offering the same interactive architecture to users who may never touch a device.
Who can access Cowork now, and what’s coming next for Windows and other platforms
For now, Cowork remains exclusive to Claude Max subscribers who use the macOS desktop app. Users in other subscription tiers – Free, Pro, Team, or Enterprise – can join the waitlist for future access.
Anthropology has indicated clear intentions to expand the scope of the feature. The blog post clearly indicates plans to add cross-device sync and bring Cowork to Windows as the company learns from a research preview.
Cherny set expectations appropriately, calling the product “early and raw, similar to how Claude Code felt when it was first released.”
To access Cowork, Max subscribers can download or update the Claude macOS app and click “Cowork” in the sidebar.
The real question facing AI adoption in enterprises
For technical decision makers, the implications of working together extend beyond a single product launch. The bottleneck to AI adoption is changing – it is no longer typical intelligence that is the limiting factor, but rather workflow integration and user trust.
Anthropic’s goal, the company says, is to make working with Claude less like operating a tool and more like delegating to a colleague. Whether ordinary users are willing to hand over access to folders to AI that might misinterpret their instructions remains an open question.
But Cowork’s development speed – a key feature built in ten days, perhaps by the company’s AI – predicts a future in which these systems’ capabilities accumulate faster than organizations can evaluate them.
The chatbot has learned how to use a file manager. What he learns to use next is anyone’s guess.
