Picture the part of your job that isn’t really your job. The monthly report that means opening eleven files and copying numbers between them. The forty candidate CVs that need to become one comparison table. The folder of meeting notes that should have been a summary two weeks ago. None of this is the work you were hired for — it’s the sediment that settles around it.
Claude Cowork is Anthropic’s answer to that sediment, and the important thing to understand is who it was built for: not developers. It was built for the person with the eleven files.
The one-line version
Cowork is a mode inside the Claude desktop app where you give Claude a folder and it does multi-step work on the real files inside it — reading them, producing new ones, reorganising what’s there — and then shows you the result for approval.
That’s the whole idea. The leap isn’t intelligence; it’s access and persistence. Chat is brilliant but blind: it can only see what you paste into it, and it stops after every answer. Cowork can see your actual working folder, and it keeps going until the job is done.
Chat, Cowork, Code — three doors, one Claude
Open the Claude desktop app and you’ll find a mode selector. It’s worth knowing what each one is for, because picking the wrong door is the most common way people conclude “AI isn’t useful for my job.”
- Chat — the conversation you already know. Ask, get an answer, ask again. Perfect for thinking out loud, drafting, and questions. It cannot reach your files.
- Cowork — you point Claude at a folder and describe an outcome. It works through the steps on the real files and comes back with finished work. Built for non-technical professionals.
- Code — Claude Code, the same agent aimed at a codebase, with the tools a software project needs: diffs, version control, running the project to check it works.
Same Claude underneath all three. What changes is what it can see and what it’s allowed to do. If you’ve been evaluating “AI for my team” purely through the Chat window, you’ve been looking at roughly a third of the product.
What “give it a folder” actually means
This is the part that makes people nervous, and it shouldn’t. Cowork runs on a folder-permission model: you explicitly grant read, write, and create access to specific directories. That folder is Claude’s whole world for the session. Files elsewhere on your machine are not visible to it, not because it’s being polite, but because it genuinely cannot see them.
So the safe way to start is obvious and worth doing: make a folder, put copies of the files in it, and point Cowork there. Once you’ve watched it work a few times, you’ll widen the scope naturally.
And for anything consequential — sending an email, finalising a document that goes outward — it stops and asks. You review, then approve. The autonomy is real, but it’s bounded, and the boundary is you.
What people actually use it for
Anthropic published usage data drawn from around 1.2 million Cowork sessions, and the shape of it is the most useful thing written about this product so far:
| What people use Cowork for | Share of sessions |
|---|---|
| Business processes & operations | 33.4% |
| Content creation & copywriting | 16.4% |
| Software development | 8.7% |
| DevOps & infrastructure | 7% |
| Research & intelligence | 6.4% |
Read that top row again. A third of all use is business process and operations — pulling updates into reports, building onboarding checklists, reconciling spreadsheets. The two leading categories together are roughly half of everything. Software development, the thing everyone assumes agentic AI is for, is under nine percent.
This is a product whose heaviest users are not engineers. That single fact should reframe how most organisations think about who gets a seat.
The shape of a real task
Concretely, here’s the difference between the two modes on the same problem.
In Chat, you’d paste in a spreadsheet, ask what’s wrong with it, get a thoughtful answer, paste in the next one, and repeat — you’re the one carrying data between steps.
In Cowork, you point it at the folder and say:
Go through every file in
/Q3-regional, reconcile the revenue columns againstmaster-ledger.xlsx, flag any row that differs by more than 2%, and write me a one-page summary of what you found with the discrepancies in a table.
Then you go and do something else. It opens each file, does the comparison, notices the ones that don’t line up, and produces the summary. You come back to finished work and a list of things to check — not to a conversation you have to keep feeding.
The difference isn’t that Claude got smarter between those two examples. It’s that in the second one it had hands.
It doesn’t need your laptop open any more
Cowork started as a desktop-only feature in January 2026. In July it expanded to the web at claude.ai and to mobile on iOS and Android, in beta — and that expansion came with a genuine architectural change, not just new screens.
Sessions now run remotely on Anthropic’s servers by default. A task keeps going after you shut your laptop. Scheduled work — “every Monday morning, pull last week’s numbers into the standing report” — runs whether or not any device of yours is switched on. You get the result when you next open the app.
That turns Cowork from a tool you operate into something closer to a colleague who works while you’re asleep. Which is also, reasonably, the point at which you should start caring about review habits: see our guide on trusting an agent’s output.
Where it plugs into the rest of your work
On its own, Cowork sees a folder. Connected, it sees your actual working life. Organisations wire it to the tools their work already lives in — Google Drive, Gmail, DocuSign, and industry systems like FactSet — and can deploy plugins that encode how their company does financial analysis, or hiring, or engineering review.
That last part is the quietly important one for teams. A plugin is how institutional knowledge — the way your firm formats a board pack, the checks your compliance team always wants — stops living in one senior person’s head and becomes something everyone’s Claude just knows. If you’re thinking about this at team scale, our team guide covers how to sequence it.
Who should be using this
If your work is mostly thinking — strategy, judgement, relationships — Cowork won’t do your job, and shouldn’t. But almost nobody’s work is only thinking. It’s thinking wrapped in a thick layer of assembling, formatting, chasing, and reconciling.
Cowork is aimed squarely at that layer. Anthropic’s own framing is that it handles the connective work around a role — the information management that spans nearly every job rather than belonging to any one of them. Which means the people who benefit most are often the ones furthest from the engineering org: operations, finance, HR, marketing, analysts, founders doing six jobs at once.
The honest test is a small one. Take the most annoying recurring task you own — the one you’d describe as “not hard, just tedious” — put its files in a folder, and describe the outcome you want. Most people’s opinion of what this technology is for changes somewhere in the first ten minutes.
If you don’t have the app yet, start with getting set up on Desktop. If you’re weighing Cowork against Claude Code specifically, that comparison is here.