Anthropic just launched Cowork, a new Claude feature that lets the app work with files on your computer. Claude did not suddenly become a completely different model. What changed is the way people can reach some of the same AI help developers have been using.

Why This Feels Different

Some of the most capable AI tools right now are still built around developer habits. Claude Code, Cursor, Codex, and Copilot can read files, make edits, run commands, and work through a task with a decent amount of independence. If you already write code, that can feel normal. You tell the tool what you want, watch what it changes, and step in when it needs direction.

For everyone else, the setup can be intimidating before the AI even gets involved. A terminal is the text-based window where developers type commands directly into a computer. An IDE is a coding app, basically a workspace for writing and managing software. Those are useful tools, but they are not where most people do their everyday work.

That matters because a lot of useful AI work is not really about programming. It is sorting through files, comparing drafts, cleaning up notes, building a simple spreadsheet, checking whether something has already been published, or turning a messy folder into something organized. The AI may be capable of helping with those jobs, but the person still needs a way to ask for that help without first learning a developer workflow.

Simon Willison wrote that Claude Code has always been closer to a general computer assistant than a coding tool alone. I think that is the right way to see it. The hard part is making that assistant feel reachable to people who do not want to open a terminal or learn how a codebase works.

Cowork is Anthropic’s attempt to do that.

What Cowork Does

Cowork launched on January 12 as a research preview in the Claude desktop app. It appears as its own tab next to Chat and Code.

The setup is pretty simple. You choose a folder on your computer and give Claude permission to work there. Claude can read the files in that folder, create new ones, and edit existing ones. Instead of dragging documents into a chat one at a time, you point Claude at the place where the work already lives.

From there, you can describe a goal in normal language. Cowork makes a plan, shows you what it is doing, and works through the steps. An “agent” is just the current word people use for an AI system that can take a task, break it into smaller parts, use tools, and keep going without needing a new instruction for every single click.

Simon’s example is a good one because it is ordinary. He gave Cowork access to his blog drafts and asked it to find drafts from the last three months, check whether any had already been published, and suggest which ones were closest to ready. Cowork searched his site, looked through the draft files, compared what it found, and came back with a prioritized list. It even caught that one draft had already appeared somewhere else.

That is not a flashy use case, but it is exactly the kind of task people avoid because it is boring and scattered. You know the answer is somewhere in your files. You know you could figure it out manually. You also know it would take a while, and you would rather be doing almost anything else.

Why the Desktop Piece Matters

A lot of AI agents being built right now are centered on the browser. ChatGPT agent mode and Gemini agent mode can move through websites, fill out forms, click buttons, and handle web-based tasks. That can be useful, especially when the work really does live on the web.

Cowork is aimed at a different place: the files already sitting on your computer. It works through the Claude macOS app, inside a workspace that you choose. Your notes, documents, spreadsheets, PDFs, and drafts can stay in a normal folder, and Claude works with them there.

That distinction sounds small, but it changes the kinds of jobs that make sense. Browser agents have to deal with websites changing layout, pages loading strangely, sign-in screens, bot checks, and all the other weirdness of the web. Cowork is closer to having someone sit down next to your folder and help you sort through it.

Anthropic also built Cowork with a stronger boundary around what Claude can touch. Under the hood, it runs in a sandboxed virtual machine using Apple’s Virtualization Framework. In plain English, that means Claude is operating inside a contained environment, not roaming around your whole computer. It can only see the folders you give it.

That does not make it risk-free, but it is more reassuring than a vague promise that the AI will only do what you asked. If a tool is going to read and change local files, the boundaries need to be understandable.

Who This Is Really For

The easiest way to underestimate Cowork is to judge it only as a power-user tool. Developers already have several ways to get AI help inside their workflows, and many of those tools are excellent. Cowork is more interesting when you imagine someone in operations, finance, marketing, HR, research, admin, or a small business trying to get through a pile of normal work.

Those teams have documents everywhere. They have folders full of exports, old reports, half-finished drafts, meeting notes, screenshots, PDFs, and spreadsheets with names like final_v3_revised_actual_final.xlsx. Their work is not less complex than coding. It is just stored in different places and wrapped in different habits.

A tool like Cowork gives those users a more believable way to use an AI agent. They do not have to learn what a command line is. They do not have to paste twenty files into a chat window and hope the model remembers them. They can start with a folder and a request:

  • Find the duplicate files and make a cleanup plan.
  • Turn these meeting notes into a short project update.
  • Compare these drafts and tell me what changed.
  • Build a spreadsheet from these receipts.
  • Pull the important dates out of these PDFs.

None of those examples require the AI to be magical. They require the tool to understand files, keep track of steps, and show enough of its work that a person can trust it.

This is also why Cowork probably matters for Anthropic’s business. Claude Code has been a strong product for technical teams, but companies are not made only of developers. If Anthropic wants Claude to become part of everyday work across an organization, the interface has to make sense to people outside engineering.

What Still Needs To Be Better

There are some obvious limits.

Cowork is only available to Claude Max subscribers right now, and that plan starts at $100 per month. That is not exactly “for everyone,” even if the interface is more approachable. Research previews often start this way, but price still matters if the goal is broader use.

Safety also matters. Giving an AI permission to read and edit files is a big deal. Even with a sandbox, mistakes can happen. The model might misunderstand an instruction, overwrite something important, or follow bad directions hidden inside a document or webpage. Anthropic has published work on prompt injection defenses, but agent safety is still an open problem.

For now, the common-sense version is simple. Give Cowork a narrow folder, not your whole digital life. Keep backups of anything important. Start with low-risk tasks. Watch what it is doing until you understand how it behaves.

That may sound cautious, but it is the right way to approach a tool at this stage. Something can be genuinely useful and still need supervision.

Where This Might Go

AI tools are starting to move out of the chat box.

Chat is familiar, and it is still useful. But a lot of real work does not happen as a clean back-and-forth conversation. It happens across folders, files, drafts, screenshots, notes, and half-remembered decisions. If AI is going to help with that work, it needs to meet people where the work already is.

Cowork is one early version of that idea. It takes capabilities that have mostly been available to technical users and wraps them in a desktop interface that is easier to explain: choose a folder, describe the job, review what Claude does.

The details will change. Other companies will build similar tools. Prices will move around. Some versions will be clumsy, and some will probably feel much better than this first research preview. But the direction feels important.

The next wave of useful AI products may not come only from models getting smarter. It may come from making the models we already have easier for normal people to use without feeling like they have to become developers first.

Cowork is not the finished version of that future. It is one early sign that the interface problem is being treated as part of the product, not as an afterthought.

Further reading: