Anthropic’s Cowork turns Claude into a file-crunching assistant for non-coders

Anthropic dropped something interesting on Monday. Cowork is a new agent capability that basically takes what <a href="https://code.allwinchina.org/ai-tools/claude-code/" title="Claude Code review”>Claude Code does for developers and strips away the command-line nonsense so normal people can use it too.

What caught my attention isn’t just the product itself — it’s that the team apparently built the whole thing in about a week and a half, mostly using Claude Code. Yes, they used their own coding tool to build a tool that makes their coding tool accessible to non-coders. That’s a neat recursive loop if you think about it.

Cowork is rolling out as a research preview for Claude Max subscribers — that’s Anthropic’s $100 to $200 per month tier — through the macOS desktop app. If you’re paying that much, you probably expect something more than a fancy chatbot.

The origin story is the best part

Anthropic launched Claude Code back in late 2024 as a terminal-based tool for developers to automate programming grunt work. It did well. But then something weird happened: people started using it for everything except coding.

Boris Cherny, an engineer at Anthropic, posted about this on X. People were using Claude Code for vacation research, building slide decks, cleaning up email, canceling subscriptions, recovering wedding photos from a hard drive, monitoring plant growth, controlling their oven. The list is genuinely bizarre and I love it.

This is the kind of shadow usage that product managers dream about. Users were contorting a developer tool into a general-purpose assistant because the underlying model — Opus 4.5 — was good enough to handle whatever they threw at it. Anthropic noticed and decided to make it official.

How Cowork actually works

Instead of a chat interface where you paste text and get text back, Cowork works with folders. You designate a specific folder on your machine that Claude can access. Within that sandbox, it reads files, modifies them, or creates new ones.

Anthropic’s examples are pretty straightforward: reorganize your messy downloads folder with intelligently renamed files, generate an expense spreadsheet from receipt screenshots, or draft a report from scattered notes across documents.

The architecture runs on what they call an “agentic loop.” When you give it a task, it doesn’t just generate a response. It formulates a plan, executes steps in parallel, checks its own work, and asks for clarification when stuck. You can queue multiple tasks and let it process them simultaneously. Anthropic describes this as feeling “much less like a back-and-forth and much more like leaving messages for a coworker.”

That’s a fair description. The async nature of it is actually smart — you don’t sit there watching it work. You dump tasks in and come back later.

The recursive build story

Here’s where it gets meta. During a livestream, Dan (Anthropic’s product lead) mentioned that Cowork was built almost entirely using Claude Code itself. The team knocked it out in about a week and a half.

This isn’t just a cute anecdote. It demonstrates something real about how AI development is changing. When your own tool is good enough to ship production features in days, the feedback loop becomes incredibly tight. The team at Anthropic is eating their own dog food and it seems to be working.

Where this fits in the bigger picture

Anthropic is positioning Cowork to compete not just with OpenAI and Google in conversational AI, but with Microsoft’s Copilot in the productivity tool space. That’s a crowded market, but Cowork has a different angle.

Most AI productivity tools work through chat interfaces or embedded assistants. Cowork works in your actual file system. It’s less about conversation and more about delegation. You tell it what to do with your files and it goes off and does it.

The folder-based sandbox is both a feature and a limitation. It gives Claude access to your data, which means you need to trust it. But it also keeps the AI contained to a specific workspace rather than roaming your entire machine. That’s a reasonable compromise for now.

What I think about this

I’ve been skeptical about AI agents for a while. Most of them promise autonomy but deliver glorified macros. Cowork looks different because it’s built on actual usage patterns rather than hypothetical use cases.

The fact that real users were already doing this stuff with Claude Code — even when it was inconvenient — suggests there’s genuine demand. Anthropic just removed the friction.

That said, $100-200 per month is steep. This is clearly aimed at power users and professionals who can justify the cost. For everyone else, it’s a preview of where things are heading rather than something to buy today.

The recursive build story is the part that sticks with me. If Anthropic can ship a product like Cowork in ten days using their own tools, the pace of improvement is only going to accelerate. That’s both exciting and a little scary.