Anthropic released Cowork on January 12th, a new feature that essentially gives non-developers the same agentic capabilities that made Claude Code popular with programmers. It's available now in the macOS desktop app for Max plan subscribers ($100 or $200/month). Everyone else can join a waitlist.
The pitch is simple
You point Claude at a folder on your computer. Claude can then read, edit, or create files in that folder. The example use cases in the announcement: sorting your downloads, creating spreadsheets from screenshots, drafting reports from scattered notes.
This isn't new technology. It's the same Claude Agent SDK infrastructure that powers Claude Code, repackaged with a friendlier interface. Anthropic is explicit about this: Cowork is "built on the very same foundations."
Why this exists
Here's the interesting part. Anthropic says developers started using Claude Code for "almost everything else" beyond coding. Lenny's Newsletter ran a piece documenting 50 non-technical workflows people had cobbled together with Claude Code, from file organization to synthesizing customer calls. So Anthropic built Cowork to lower the barrier.
The framing in the announcement is telling: "much more like leaving messages for a coworker." You set a task, Claude makes a plan, works through it, loops you in. You can queue up multiple tasks and let them run in parallel.
The safety asterisks
Anthropic buried the important stuff in the middle of the blog post and in a separate help center article. Some highlights:
Claude can take "potentially destructive actions (such as deleting local files)" if you're not careful with your instructions. The recommendation: give Claude "very clear guidance around things like this." Not exactly reassuring.
Prompt injection remains a real concern. Anthropic's own research shows attack success rates around 1% against their current defenses. That sounds small until you remember we're talking about an agent that can modify files on your computer and browse the web on your behalf.
If you enable Claude in Chrome as a connector (which Cowork supports), the attack surface expands considerably. Every webpage, email, and document becomes a potential vector.
The practical limitations
This is a research preview. That means:
No Windows support yet. macOS only, desktop app required. Cross-device sync isn't available. And you need the Max plan, which starts at $100/month.
Anthropic says they'll "improve it rapidly from here" and add Windows eventually. But right now, this is very much an early access product aimed at power users who are already paying premium prices.
What Cowork can actually do
The announcement mentions connecting to existing "connectors" (Anthropic's term for integrations with external services) and a set of "skills" for creating documents, presentations, and other files. If you pair it with the Chrome extension, Claude can handle tasks that need browser access.
The comparison to Claude Code makes sense once you understand the underlying architecture. The Agent SDK gives you primitives for a feedback loop: gather context, take action, verify work, repeat. Cowork just wraps this in a more accessible interface.
So should you care?
If you're already a Max subscriber and you've been manually orchestrating Claude for file-based tasks, Cowork is probably worth trying. The research preview label means expectations should be calibrated accordingly.
If you're on a lower tier, the waitlist is your only option. No word on when or if this reaches Pro subscribers.
The broader signal here is that Anthropic sees agentic workflows as the future of their consumer product, not just developer tools. They're watching how people use Claude Code and shipping features to meet that demand. Whether the safety measures are adequate for mainstream users is a question Anthropic is explicitly deferring: "We're releasing Cowork early because we want to learn."
Windows support is supposed to come eventually.
