Claude Cowork is Anthropic's desktop agent for knowledge work. You give it a goal, point it at a folder on your own computer, and it goes off and does the task, then hands back a finished file. No terminal. No code. That is the short answer to what is Claude Cowork, and most of this article is just the fine print on that sentence.
The longer version matters, because "agent that touches your files" can mean a lot of things, and a few of them are scary.
What is Claude Cowork, and what does it do
Here is the shape of a real Cowork job. You have forty receipts sitting in your Downloads folder as PDFs and screenshots. You tell Claude to pull the vendor, date, and amount out of each one and build you a spreadsheet. It reads the files, extracts the numbers, writes an .xlsx, and tells you when it is done. You never opened a single receipt yourself.
Anthropic frames the whole product around that ending: you describe an outcome, and Claude works on your computer, your local files, and your apps to return a finished deliverable. It is good at a few kinds of grunt work in particular. Cleaning up messy folders. Building documents out of source material. Reading across a pile of files and pulling the through-line into one place. Yanking structured data out of unstructured stuff like PDFs and receipts. None of that is hard, exactly. It is just slow when a human does it.
That is the difference from a normal chat. A chat tells you how to clean up your Downloads folder. Cowork cleans it up.
It can also run on a schedule and check in with you as it goes, so a weekly report that always pulls from the same five files becomes a thing you set once.
Who Claude Cowork is for
This is the part Anthropic is unusually direct about. Cowork is for people whose job involves work that eats hours but is not technically hard: researchers, analysts, operations teams, legal, finance. The official line is "anyone whose workday includes tasks that are time-consuming but not technically complex."
If you have ever renamed 200 files by hand, or copied numbers out of PDFs into a sheet one row at a time, or stitched six documents into one report at 6pm because it was due in the morning, you are the target user. You do not need to know what a terminal is. That is the entire point.
How Claude Cowork relates to Claude Code and the Claude apps
A quick map, because the names blur together.
Claude Code is the agent that lives in your terminal, and it was built for developers. It is powerful. What surprised Anthropic was that non-technical teams inside the company started using it anyway for non-coding work, mining data and building little throwaway tools, because it could just do the job instead of describing it. Cowork came out of watching that happen. It takes the same agent engine that powers Claude Code and puts it inside the Claude desktop app, with no command line in sight.
So the relationship is plain. Same engine underneath. Different front door. Claude Code asks you to be comfortable in a terminal; the Claude Cowork app does not ask you for anything except a folder and a goal.
The Claude apps you already know (the chat window on web, desktop, and mobile) are the conversational side. Cowork is the part of the desktop app that goes and does multi-step work on your files instead of just talking.
The part most explainers skip: skills
This is where Claude Cowork gets more useful than it first looks, and it is the reason a non-coder should care.
Cowork runs on skills. A skill is a small folder with a file called SKILL.md inside it, plus any helper files, and it teaches the agent how to do one specific job your way. Anthropic ships Cowork with built-in office skills for the formats everyone actually uses: Excel, PowerPoint, Word, and PDF. That is why "make me a branded report" produces a real .docx and not a wall of text.
The SKILL.md format is an open standard, not a Cowork-only thing. The same skill file runs in Claude Code, in the Claude apps, and in Cowork. Write it once, it works in all of them. So a skill is portable in a way most "automations" are not, and you are not locked into one tool.
Which raises the obvious question: who writes the SKILL.md?
You can, without touching code. That is what Knack is for. Knack turns a short interview about your process into a finished SKILL.md folder in the open Anthropic format, the exact same format Cowork reads. You answer questions about how you do a task. Knack writes the skill. Then Cowork runs it. A non-technical person can build the very thing the agent then executes, which is a strange and good place to be standing in 2026. There is more on the mechanics in how Claude Cowork uses skills and on the format itself in what a Claude skill is.
Is it safe to let it touch your files
Fair worry. The design answers it in two ways.
First, scope. Cowork only sees the folders you hand it. You point it at Downloads or a specific project directory, and that is its world. It cannot wander your whole drive.
Second, the sandbox. Each task runs inside an isolated virtual machine on your own computer, so the work is walled off from the rest of your system even while Claude makes real changes to the files you granted. You choose between "ask before acting" and "act without asking," and deleting anything permanently always requires your explicit yes. The desktop app has to stay open while a job runs, which is a small annoyance and also a useful reminder that something is actively working.
None of this makes it magic. An agent acting on real files can still make a real mess if you give it a vague goal and full autonomy. Start it on a copy of a folder for the first few runs. Watch what it does.
Platforms and availability (with the honest caveats)
Claude Cowork runs in the Claude desktop app on macOS and Windows. The desktop app is required; there is a phone-pairing beta that lets you fire off tasks from mobile, but the work itself happens on your computer.
On timing and plans, here is what Anthropic's own release notes and getting-started guide say. Cowork started as a research preview on January 12, 2026, macOS only, for Max plans, and widened to Pro a few days later. It reached general availability on April 9, 2026, on both macOS and Windows. The support docs now list it as available on the paid plans: Pro, Max, Team, and Enterprise.
One caution worth flagging plainly: the exact tier access has shifted between the preview and now, and different Anthropic pages word it slightly differently. If your plan's access matters to you, check the Cowork product page and your own account before you count on it. Plan boundaries are the one detail in this article I would not tattoo on.
Where to go next
If you want to actually drive it, the step-by-step walkthrough covers a first real task end to end. If you want it to do your job and not a generic one, the move is to build a skill for your specific process first. A non-coder can do that in an afternoon with Knack, ship the SKILL.md, and let Cowork run it on Monday.