Claude Cowork: What It Is and How to Use It
Since 16 September 2026, Cowork has been part of Claude’s ordinary chat on Pro and Max rather than a separate mode, and Team and Free are due to follow. We’ve updated this guide to match, and there’s more on agents in our guide to AI agents.
Cowork is the mode where Claude does the work rather than telling you how to do it. You hand over a whole task (“go through these forty support emails, group them by issue, draft a reply to each”) and it reads your files, makes a plan, works through the steps, and comes back with finished output while you do something else. Since 16 September 2026 it’s no longer a separate mode on Pro and Max: chat and Cowork are one Claude, and Claude decides itself whether a request needs a quick answer or a task.
The work runs on Anthropic's servers rather than your machine, so you can start a task, shut your laptop and leave. It reaches you on the web, on desktop and on your phone, and scheduled jobs run with nothing of yours switched on at all. The shift is from an assistant that needs you sitting there to a colleague you can walk away from.
What Cowork actually is
If you have only used Claude as a chat box (ask, get an answer, copy it out), Cowork is a different way of working with the same Claude. Instead of having a conversation, you delegate a job.
The practical difference is who does the assembling. In a chat, you gather the inputs, paste them in, read the reply, and stitch the result together yourself. In Cowork you describe the outcome, and it does the gathering, the steps and the stitching. You come back to something finished.
Cowork vs Claude Code vs the chat box
This is the comparison people search for most, and the confusion is understandable: all three are Claude, and the names do not tell you who each is for.
| Who it’s for | What you give it | What you get back | |
|---|---|---|---|
| Claude chat | Everyone | A question, or something to work on now | An answer you read and use immediately |
| Cowork | Anyone with recurring multi-step work, no technical skill needed | A whole task, and access to the files and tools it needs | Finished output, produced while you do something else |
| Claude Code | Developers, and increasingly technical non-developers | A codebase and a goal | Working software: written, run, tested and fixed |
On Pro and Max, the first two rows now happen in the same place: you ask as usual, and Claude picks the row for you. On Team and Enterprise, Cowork is still its own mode for now.
The short version: chat is for thinking, Cowork is for delegating, Claude Code is for building. Most non-technical teams will get the most from Cowork and never need to open Claude Code, though the boundary is softer than it used to be, and plenty of non-developers now use Code to build small internal tools.
Is it free, and where does it run?
Cowork is part of Claude's paid plans rather than something you buy separately, and there is no free-standing free tier for it. The free Claude account gets the chat experience.
On Pro and Max, the merged Claude has been rolling out gradually on the web, desktop and mobile since 16 September 2026. Anthropic says Team and Free will follow soon, and Enterprise admins get at least 30 days’ notice before anything changes. Until your account moves over, Cowork is in the desktop app on every paid plan, and on the web and mobile in beta on Pro, Max and Team, and on Enterprise if it’s been switched on. Usage limits have been moving as Anthropic expands the rollout, so check the current plan comparison rather than trusting any article on that specific point, including this one.
Where it runs is simpler and more useful to know: the work happens on Anthropic's servers, and you reach it from the web, the desktop app on Mac or Windows, or your phone. Because the session belongs to your account rather than to a machine, you can start something at your desk, check it from your phone in a meeting, nudge it (“actually, leave out the Q3 numbers”) and collect the result from whatever device you happen to have later.
Three details are worth knowing. Anything that works with files on your own computer, uses your browser or controls your computer needs the desktop app running. On Enterprise, cloud sessions are off by default until an admin switches them on. And on Pro and Max the “Only on your computer” option is removed on 6 October 2026, so new tasks run in the cloud.
Scheduled tasks run with nothing of yours turned on. A recurring job that needs your laptop awake at six in the morning is a job that quietly fails half the time. One that runs on a server actually happens, which is the difference between a clever demo and something you come to rely on.
How to use it: your first three tasks
The mistake almost everyone makes is starting with something ambitious and important. Start instead with something recurring and low-stakes, so that when it gets something wrong, as the first attempt usually does, nothing breaks and you learn how to brief it better.
- Something you do every week that you resent. Compiling a report from the same four places. Summarising a recurring meeting. Formatting the same kind of document. Recurring is what matters, because you want to run it more than once so you can improve the brief.
- Something with a pile of inputs and a clear output. Forty emails into five themes. A folder of documents into one summary. This is where it comfortably beats doing it yourself, and where the result is obviously right or obviously wrong.
- Something scheduled, once the first two work. A Monday briefing built from last week's threads and this week's calendar, sitting ready before you open your laptop. This is the one that changes how the week feels.
Brief it the way you would brief a capable new colleague who has never seen your company: what the output should look like, what good means, what to leave out, and what to do if something is ambiguous. The instinct to write a short prompt is worth resisting here. A task you describe in three sentences comes back needing three rounds of correction.
Why “you can walk away” is the whole point
Think about what makes delegating to a person valuable. Part of it is their competence, but the bigger part is that once you have handed something over, you get your attention back. You do not stand over their shoulder for an hour.
AI that runs on your own device never quite had that property. Technically you could start something and look away, but you were tethered: you could not close the laptop or go into a meeting without risking the work stopping. So most people babysat it, which is exhausting and quietly caps how much anyone is willing to hand over.
Cutting that tether is the actual upgrade. Once walking away is genuinely safe, the tasks people are willing to delegate get bigger, because the cost of delegating (sitting there watching) drops to nothing.
What it is and isn't suited to
Being honest about the second half of this saves teams a lot of disappointment.
It suits work that is multi-step, involves gathering from several places, has a recognisable finished shape, and recurs often enough that improving the brief pays off. The low-glamour work that eats mornings.
It does not suit anything where you cannot describe what good looks like before it starts, anything requiring judgement that depends on context nobody has written down, or anything where being wrong is expensive and nobody will check the output. That last one deserves emphasis. The usual failure mode is something confident and wrong that nobody notices because it looked finished, rather than a refusal.
Review the first several runs of any task properly. Once you have seen it handle the same job right a few times, you can relax, and by then you will have a real sense of where its judgement is reliable and where it is not.
Working out which of your team's recurring tasks are worth handing over, and writing briefs good enough that the output comes back usable, is most of what a Deployed Kickstart half-day covers. The Partner programme is where teams go from one working task to a dozen.
Frequently asked questions
What is Claude Cowork?
The part of Claude that does the work rather than telling you how to do it. Instead of having a conversation you delegate a whole task (go through these forty emails, group them by issue, draft a reply to each), and it reads your files, plans the steps, works through them and returns finished output while you do something else. Since 16 September 2026 it is no longer a separate mode on Pro and Max: chat and Cowork are one Claude, and Claude decides from your request whether it needs a quick answer or a task.
What is the difference between Claude Cowork and Claude Code?
Cowork is for anyone with recurring multi-step work and needs no technical skill: you give it a task plus access to the files and tools it needs, and get finished output back. Claude Code is for working on software: you give it a codebase and a goal, and it writes, runs, tests and fixes code. The short version is that chat is for thinking, Cowork is for delegating, and Claude Code is for building.
Is Claude Cowork free?
No. Cowork is part of Claude’s paid plans rather than a separate purchase, and a free account gets the chat experience instead. Which paid plans include it, and at what usage limits, has been moving as the rollout expands, so check the current plan comparison rather than relying on any article for that specific detail.
Does Claude Cowork work on Windows and mobile?
Yes. Cowork is in the desktop app for Mac and Windows on every paid plan. On Pro and Max it has been part of the merged Claude on web, desktop and mobile since 16 September 2026, rolling out gradually. Until an account moves over, web and mobile are in beta on Pro, Max and Team, and available on Enterprise if it has been switched on. Because the work runs on Anthropic’s servers and the session belongs to your account rather than a machine, you can start a task at your desk, check it from your phone, nudge it mid-run, and collect the result from any device later.
Does Cowork keep running when I close my laptop?
Yes, and this is the main point of it. The work happens on Anthropic’s servers rather than your machine, so you can start a task and leave. Scheduled tasks go further: they run with nothing of yours switched on, which is the difference between a recurring job that quietly fails half the time and one that actually happens. The exception is a task that works with files on your own computer, uses your browser or controls your computer, which needs the desktop app running.
What should my first Cowork task be?
Something recurring and low-stakes rather than something ambitious and important, because the first attempt usually gets something wrong, and you want that to be harmless. Good candidates are a weekly job you resent, or a pile of inputs with a clear output such as forty emails into five themes. Once those work, move to a scheduled task.
What is Claude Cowork not good at?
Anything where you cannot describe what good looks like before it starts, anything needing judgement that depends on context nobody has written down, and anything where being wrong is expensive and nobody will check. The failure mode to watch for is confident, wrong output that nobody catches because it looked finished. Review the first several runs of any task properly.
How should I brief a Cowork task?
The way you would brief a capable new colleague who has never seen your company: what the output should look like, what good means, what to leave out, and what to do when something is ambiguous. Resist the instinct to write a short prompt: a task described in three sentences comes back needing three rounds of correction.
Found this useful? Send it to someone who needs it.