How to Use Claude Cowork: Setup, Tasks, and Real Limits
Most guides to Claude Cowork start with “type your task and press enter,” which is true and almost useless. The part that decides whether Cowork works for you is a choice you make before typing anything: which surface you open it in. Desktop, web, and mobile are not three views of one product. They have different reach, and picking wrong is why people conclude the tool is broken.
This walks through the setup that matters, the three permission modes, how scheduled tasks actually behave, and the specific point where Cowork stops. Every detail here was checked against Anthropic’s own documentation on August 7, 2026.
Before anything: you need a paid plan
Cowork is available on Pro, Max, Team, and Enterprise. The Free tier does not include it, and there is no separate purchase or metered add-on. Your plan sets your usage limits and nothing else.
That matters more than it sounds, because Anthropic states plainly that Cowork “consumes more of your usage allocation than chatting.” The company’s own FAQ is blunt about why: agentic tasks coordinate multiple sub-agents and tool calls, so a single Cowork run can draw down what a dozen chat messages would. Pro at $17 a month on annual billing, or $20 monthly, is the cheapest way in, and Anthropic’s pricing page attaches a warning to it recommending an upgrade for heavy use. Max 5x is $100 and Max 20x is $200. Our Claude Cowork pricing breakdown covers how the limits behave once you are actually working.
Step one: pick the surface, because they are not equivalent
This is the decision that determines what Cowork can do for you, and it is the thing most tutorials skip.
| Surface | Where it executes | What it reaches |
|---|---|---|
| Desktop (macOS, Windows) | On your machine | Local folders, local apps, your browser, your screen |
| Web (claude.ai) | Anthropic’s servers | Connectors and the web only |
| Mobile (iOS, Android) | Anthropic’s servers | Connectors and the web only |
Anthropic’s own answer to “do I need the desktop app?” is the clearest statement of the split: “The desktop app adds what web and mobile can’t reach: the folders and applications on your own computer.”
So the rule is simple. If the job involves files sitting in a folder on your laptop, you need desktop. If you want the job to run while you are asleep, you want the remote surfaces, which are still in beta. A workflow built around local file paths does not transfer to the cloud version, and discovering that after building it is a genuinely annoying afternoon.
One caveat on the “close your laptop” promise. Sessions keep running with the desktop app closed, but Anthropic’s getting-started guide adds a condition: “If your task uses local files, your browser, or your computer, keep the desktop app open.” The unattended promise applies to work that lives in connectors and the web, not work that lives on your hard drive.
Step two: start a task
In the message box, select Cowork. It sits next to Chat rather than in a separate app.
Then describe an outcome instead of a prompt. This is the actual behavioural change and it takes a week to get used to. “Turn these 40 spreadsheets into one reconciled summary and flag anything that moved more than 10%” is a Cowork request. “How do I reconcile spreadsheets” is a Chat request. Cowork plans the work, splits large projects into chunks that run in parallel, and shows you each step as it goes so you can redirect mid-run.
Review the plan before authorising it. Correcting a wrong assumption in the plan takes ten seconds; discovering it three hours into a run does not.
Step three: set the permission mode deliberately
Cowork has three approval levels, and the default is the safe one. Most people should change it exactly once, on purpose, and then leave it alone.
- Manual pauses and asks for approval on every action. Correct for your first week and for anything touching client data.
- Auto approves actions after a safety review. Pro and Max only.
- Skip runs with no pauses and no checks.
Deleting files is the exception to all three. Anthropic requires explicit permission before Claude permanently deletes anything, regardless of mode.
Access is scoped by you rather than assumed. Claude reads and writes only in the folders you connect, and connector access is managed from the + menu or under Customize then Connectors.
Step four: give it standing instructions
Under Settings, then Cowork, there is an Edit control next to Global instructions. Put your standing preferences there once: how you want documents formatted, which sources you trust, what to never touch.
Projects do the same job with a narrower scope. A Project is a workspace with its own files, context, instructions, and memory, so related tasks inherit shared context instead of starting cold each time. Put your style guide, account list, and standing rules in a Project and every task inside it starts already briefed.
One gap worth knowing before you rely on it: memory from your Chat conversations does not carry into Cowork. Projects are the only shared-context mechanism, so anything you told Chat about how you work has to be told again.
Step five: schedule the ones you will actually read
Type /schedule in any Cowork task, or click Scheduled in the sidebar. Scheduled tasks run in the cloud, which means they do not need your computer awake.
The budgeting trap here is real and catches people in month two. A scheduled task consumes your usage allowance every time it fires, whether or not you open the output. A daily briefing you skim twice a week is a standing draw on the allowance that funds the work you actually care about. Schedule the ones you read.
What computer use does, and when it kicks in
Cowork can drive a real screen, opening applications and navigating a browser for tasks with no clean API. Anthropic brands this Dispatch, and the escalation order is the useful part:
“Claude reaches for your connectors and integrations first, falls back to your browser when needed, and only uses your screen as a last resort.”
That ordering is deliberate, because a direct integration is faster and more precise than clicking through a UI. When Claude does take the screen it asks permission per application, shows you what it is doing, and can be stopped at any step. Computer use is in research preview for both Cowork and Claude Code.
The practical read: screen control is the escape hatch for the long tail, not the main event. If you find Cowork constantly resorting to it, the fix is usually a connector you have not set up.
Where Cowork stops
Cowork’s triggers are a schedule or you asking. That is the complete list, and it is the ceiling on what you can hand off.
The unattended-work gap closed in 2026: ChatGPT Scheduled Tasks and Claude Cowork both run on their own. The remaining gap is what starts them. These run on a timer, not in response to an event in your apps, so "when this happens, do that" is still outside what they do.
The gap shows up the moment your work is reactive rather than planned. A new lead emailing you, a contract coming back signed, an invoice hitting 30 days overdue: none of those start anything. You still have to notice, then open Cowork and ask. For work that begins when something happens in one of your other systems, the schedule is the wrong instrument, and running a task every fifteen minutes to simulate an event trigger burns allowance for the privilege of still being late.
There is a second limit that surprises teams. Cowork sessions are not shared between users, so a Cowork task is something you run rather than something your colleagues can hand work to. Anthropic has moved partway toward addressability with Claude Tag, where @Claude is a taggable Slack teammate, though that runs on Team and Enterprise plans and a colleague tagging Claude is still a person asking.
Pairing it with something that reacts
Carly covers the trigger side rather than replacing the Cowork side. It gets its own email address so people and systems can reach it directly, and it fires on inbound events across 260+ integrations instead of waiting for a clock. Pricing is free Zapier-style workflows, with AI agents from $35/month.
The two compose better than they compete, and you do not have to choose between them. Carly runs an MCP server at https://carlyassistant.com/mcp/, which you add under Customize then connectors. Once it is there, a Cowork run can call Carly’s email, calendar, contacts, CRM, workflow, and booking tools directly. Cowork does the multi-hour thinking; Carly notices the thing that should start it and completes the action across the rest of your stack. To check that against your own setup, book a call with the team.
For the wider view, what is Claude Cowork covers the product itself, Claude Cowork limitations goes deeper on the constraints, and Claude Cowork alternatives covers the field.
FAQ
How do I turn on Claude Cowork?
There is nothing to turn on. Open Claude on desktop, web, or mobile and select Cowork in the message box, next to Chat. It requires a paid plan, so if you do not see it, check that you are on Pro, Max, Team, or Enterprise rather than Free.
Do I need the desktop app to use Claude Cowork?
Only if your work involves local files, your own browser, or your screen. Web and mobile handle anything that lives in connectors or on the internet, and they run remotely so tasks continue with your laptop closed. Desktop is the only surface that reaches your own machine.
How do I schedule a recurring task in Claude Cowork?
Type /schedule inside any Cowork task, or use the Scheduled section in the sidebar. Scheduled runs execute in the cloud with no device of yours online. Remember that each run consumes usage allowance whether or not you read the output.
Why does Claude Cowork use up my limits so fast?
Because agentic work is not chat. Anthropic confirms Cowork coordinates multiple sub-agents and tool calls per task, so one run costs far more than one message. Heavy or heavily scheduled use is the main reason people move from Pro to Max.
Can Claude Cowork start a task on its own?
Only on a schedule you set. It has no event triggers, so nothing begins because an email arrived, a record changed, or a deadline passed. Every other run starts with you asking.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
