Anthropic has released Cowork, a desktop agent that gives non-technical users the same kind of autonomous workflow that developers got with Claude Code. Instead of typing commands in a terminal, you point Claude at a folder on your Mac and let it read, modify, and create files there. Practical examples include turning a pile of receipt screenshots into a spreadsheet, drafting a report from scattered notes, or cleaning up a messy downloads folder by sorting and renaming files. It's available now as a research preview, but only to Claude Max subscribers ($100–$200/month) on the macOS app.
The origin story matters because it explains the design. After Claude Code launched, Anthropic noticed people bending a coding tool to do non-coding work — vacation research, slide decks, email cleanup, even recovering photos off a hard drive. Cowork is essentially Claude Code with the command-line complexity stripped away, built on the same Claude Agent SDK and powered by Opus 4.5. That lineage means it inherits a mature agentic loop: it plans, runs steps in parallel, checks its own output, and asks for clarification when stuck. Anthropic frames the experience as leaving tasks for a coworker rather than chatting back and forth.
Cowork also plugs into Anthropic's broader ecosystem. It works with existing connectors (Asana, Notion, PayPal, and others you've already configured in Claude), pairs with the Claude-in-Chrome extension for web automation, and supports a set of "skills" for producing documents and presentations. Anthropic highlights a built-in VM for isolation and out-of-the-box browser automation as part of the package.
The honest part of the launch is the warnings. An agent that can organize your files can also delete them if it misreads an instruction, so Anthropic urges very clear guidance around anything destructive. It also flags prompt-injection attacks — hidden instructions embedded in web content that could hijack the agent — and admits that securing real-world agent actions is still an unsolved industry problem. The takeaway for builders: treat Cowork's folder as a sandbox, start with non-critical data, and keep backups before you delegate anything irreversible.
There's a notable detail on velocity: Anthropic staff say the team built Cowork in roughly ten days, reportedly leaning heavily on Claude Code itself. Whether or not you read that as a recursive AI-builds-AI milestone, it signals how fast these capabilities are compounding — likely faster than most organizations can evaluate and govern them.
What you can do now: if you're on Max, update the macOS app and find Cowork in the sidebar; everyone else can join a waitlist, with Windows support and cross-device sync promised later. The strategic read is that the bottleneck has shifted from model intelligence to workflow integration and trust. The product is, by Anthropic's own admission, early and raw — but it's a concrete preview of delegating file-level work to an agent, and a good prompt to think hard about your access boundaries before you grant them.