Chat answers a question. Claude Cowork takes on the task and brings back something finished for you to review. That difference changes what you have to be good at.
Claude Cowork sits next to Chat in the Claude apps and uses the same agentic architecture as Claude Code without needing a terminal. You give it a goal, it plans, breaks the work into steps, runs code in an isolated environment, and hands back files. It connects to folders you attach and to apps through connectors like Google Drive, Gmail, Slack and Microsoft 365. It’s available on Pro, Max, Team and Enterprise plans, not on the free plan. Three permission modes control when it stops to ask you. The honest limits are worth reading before you start: Anthropic itself advises against pointing it at sensitive local files, and warns that isolation limits where Claude’s code runs, not what Claude reads or does.
Picture the last time you used an AI chat window to do something genuinely useful for a campaign. You probably pasted in a brief, got a draft, pasted it into a doc, went back for a second version, copied a bit of that, opened the spreadsheet yourself, and did the actual assembling by hand. The model did the thinking. You did the carrying.
Claude Cowork removes the carrying. You describe an outcome, and it works across the files and tools you’ve given it until there’s a deliverable to look at.
Anthropic’s own description is that Cowork “uses the same agentic architecture that powers Claude Code,” just without needing a terminal, so instead of responding to prompts one at a time, Claude takes on complex, multi-step tasks and carries them out for you.[1] Its product page puts the same idea more plainly: this is where you hand Claude real work.[2]
Mechanically, a task runs like this: it analyses what you asked for and makes a plan, breaks the work into subtasks where that helps, runs code and shell commands in an isolated environment on Anthropic’s servers, runs several workstreams in parallel if that suits the job, and delivers the outputs into your session to preview and download.[1]
Naming note: Microsoft ships a separate product called Copilot Cowork, in general release since 16 June 2026 after three months in Microsoft’s Frontier preview. Confusingly, “separate company” doesn’t mean “different engine underneath.”[3,11,12] Say the full name every time regardless, because the real differences are worth knowing precisely.
| Claude Cowork | Copilot Cowork | |
|---|---|---|
| What it actually does | You hand over a task, it works through it, you get files back to review | Same idea: you describe a task, it works through it and returns a result, pausing for approval on sensitive steps |
| What’s doing the thinking | Anthropic’s Claude models, always | By default, also Anthropic’s Claude models (Opus and Sonnet), with OpenAI’s GPT available in some programs and a Microsoft-trained model on the way |
| Where your files live | Folders on your own machine, plus connected apps | Nowhere local. It runs in the cloud, grounded in whatever your organisation already holds in SharePoint, Outlook and Teams |
| How you pay | One flat fee, included in your Claude plan | A Microsoft 365 Copilot licence first, then usage-based credits charged per task on top |
| The name | Both products picked up the word “Cowork” within a few months of each other in 2026. That doesn’t make them the same product wearing two logos | |
Built from Anthropic’s and Microsoft’s own product documentation, checked 25 August 2026. The model row is the one worth sitting with: these aren’t just similarly named, in Copilot Cowork’s case they can be running the same underlying AI.
Hand over the outcome, not the question. If you can’t describe what finished looks like, it isn’t ready to hand over.
That rule is doing more work than it looks like. The people I see get least out of this are the ones who bring their chat habits across, ask a question, get an answer, and conclude it’s the same product with extra steps. It is a different job. Chat is you thinking with help. Claude Cowork is you delegating.
| The job | Reach for | Why |
|---|---|---|
| “What’s a good subject line for this?” | Chat | One question, one answer, you’re iterating in real time |
| “Explain what this contract clause means” | Chat | You want to understand it, not produce anything |
| “Turn these 40 survey exports into one themed summary with a chart” | Claude Cowork | Multi-step, file-heavy, and the output is a file rather than a message |
| “Rename and sort this folder of 300 campaign assets by date and channel” | Claude Cowork | You’d never do this by copying answers out of a chat window |
| “Draft the monthly report from last month’s numbers and the team’s notes” | Claude Cowork | It pulls from several places and hands back a final document |
| “Something quick, and I’m nearly out of usage” | Chat | Anthropic states agentic tasks consume more of your allocation than chatting |
A decision rule for choosing between the two modes. The usage point is Anthropic’s own: it publishes no numeric limit for Claude Cowork, only that multi-step work is more compute-intensive and consumes more allocation.
It is not free, and it isn’t on the free plan at all. Cowork is available on paid Claude plans including Pro, Max, Team and Enterprise.[2] Pricing changes, so check Anthropic’s current pricing page before buying rather than relying on a number in this article.
On maturity, get this right because most of the coverage still circulating is out of date. Claude Cowork launched as a research preview in January 2026 on Mac only, for the top plan. It went generally available on all paid plans in April 2026, on macOS and Windows.[4] Web and mobile are in beta. Computer use, where Claude clicks and types on your actual screen, is still a research preview.[5] So it’s a released product with some beta surfaces inside it, not an experiment.
Two kinds of access, and they behave differently enough that it’s worth separating them in your head.
You attach one or more workspace folders to a session, and Claude can then read, create and modify files anywhere inside those folders, and run code against them in the sandbox.[6] The boundary is exactly as tight and exactly as loose as that sounds: Claude can only touch files in the folders you’ve connected, and outbound network access follows whatever egress settings you’ve configured.
The sentence I’d put on a sticky note, from Anthropic’s third-party deployment documentation, is the honest version: “Within an attached read/write folder, the agent can read and write every file the user’s OS account can reach.”
Which is why the single most useful setup decision has nothing to do with prompting.
Give it a folder, not your drive. Make the new folder before you start the task.
That isn’t me being cautious on Anthropic’s behalf. It’s their own advice: create a dedicated working folder for Claude rather than granting broad access, and keep backups of anything that matters.[5]
Connectors let Claude reach your apps and services, retrieve data and take actions inside them. The permission model is the reassuring part: Claude inherits each person’s own permissions from the connected service, so if you can’t reach a file, channel or record yourself, the connector can’t reach it on your behalf either.[7] Anthropic names first-party integrations with Google Drive, Gmail, Google Calendar, GitHub, Slack and Microsoft 365, alongside a public directory of several hundred more.[8]
One thing in that directory deserves reading twice, because it’s the kind of caveat most vendors bury. Anthropic is explicit that verification is not a security audit, and that once you’ve connected a community connector, it has exactly the same capabilities and access as any connector you grant. A badge tells you how much review something received. It doesn’t change what the thing can do once you’ve said yes.
| Folders on your machine | Apps via connectors | |
|---|---|---|
| What it reaches | Everything inside the folders you attach | Only what your own login can already reach in that app |
| Who controls it | You, at the moment you attach the folder | You, plus an admin on Team and Enterprise plans who enables it first |
| The risk to think about | Attaching a parent folder without noticing what’s underneath it | Approving a community connector because it looked official |
| The decision to make first | Create a fresh working folder and copy in only what the task needs | Start with first-party connectors. Add others one at a time, for a reason |
| Worth knowing | Cloud sessions reach your machine only while the desktop app is open | Custom remote connectors are reached from Anthropic’s cloud, not your machine |
Built from Anthropic’s Cowork architecture and connector documentation, checked 25 August 2026. The connectors gallery has no Claude Cowork filter, so don’t assume a given connector is supported in a Cowork session until you’ve seen it work.
A common mistake on day one is picking something impressive. Don’t. Pick something boring that you already know the right answer to, because the most important thing you’re testing in week one is whether you can tell when it’s wrong.
Before the five days, one quick filter. A good first Cowork task is recurring, file-heavy, easy to grade, low-consequence if the first attempt is imperfect, contained inside a small folder, and something you already know how to do by hand. If a candidate task fails two or more of those, park it and pick something duller.
Make a folder called claude-workspace. Copy in one month of a report you already produce by hand, plus last month’s completed version. Attach only that folder. Ask it to produce this month’s version in the same format, and compare the two yourself. You already know what good looks like, which is the point.
Repeat the same task and watch what it does rather than what it produces. You’re learning its habits: where it asks, where it assumes, what it does when a number is missing.
Add one connector, and make it a read-heavy one like Google Drive or Microsoft 365. Ask for something that pulls from the app and lands as a file in your folder. Don’t let it write back into the app yet.
Give it a genuinely messy job. A folder of unnamed exports, inconsistent survey free text, four spreadsheets that should have been one. This is where the difference from chat becomes obvious, because the work is carrying rather than thinking.
Take the task from day 1 and schedule it. Anthropic supports scheduled recurring tasks on all paid plans, running hourly, daily, weekly, on weekdays, or manually, and they run remotely even when your computer is asleep.[9] Then read the section below before you leave anything running unattended.
A five-day onboarding sequence built around one principle: only delegate work whose output you can already grade. Framework, not measured data.
One practical constraint on that last day, which catches people. Scheduled tasks work with your connectors and files saved to your Claude account, and Anthropic states they “can’t be tied to a folder on your computer.” If a task needs local files, it only runs locally. So the report you built on day 1 out of a local folder isn’t automatically the thing you can schedule, and finding that out on day 5 is better than finding it out in month three.
The hard part was never learning where the Cowork button is. It’s learning how to scope a task, choose the right permission mode, recognise when Claude has drifted off track, and decide what’s safe to hand over in the first place. Reading a support article gets you the button. Doing it, ideally with someone checking your scoping before you press go, is what gets you the judgement.
If week one goes well and you want the version that treats it as an operator rather than an assistant, we’ve written that up separately in the guide to turning Claude Cowork into a real AI operator. This piece is deliberately the front door.
What I like about Anthropic’s documentation here is that the cautions are theirs, not mine. That matters: a vendor documenting its own limits is worth paying attention to, especially when the warning is this explicit.
Start with permissions, because this is the control you’ll actually use. There are three modes governing when Claude stops to ask before acting.
| Mode | What happens | Use it for | Don’t use it for |
|---|---|---|---|
| Manual | Claude pauses and asks for approval on actions. You choose Allow or Deny each time | Anything touching a live system, anything you’d have to undo publicly, and your first fortnight | A 200-file rename, where you’ll click Allow 200 times and stop reading |
| Auto | Claude keeps working, reviewing each action for safety and blocking anything it judges unsafe, asking you when it can’t find a safer route | Bulk file work inside your own dedicated folder, research assembly, format conversion | Sending anything, spending anything, or writing into a shared system |
| Skip | No pausing, and nothing checks the actions automatically | Anthropic’s own guidance is to use it only when you completely trust every action, connector, file and app involved | Most normal business work. Treat Skip as a deliberate exception, not a default |
The three modes as documented by Anthropic, checked 25 August 2026. Deletion sits outside all three: Anthropic states Claude always asks before permanently deleting files, in any mode.
Anthropic’s own hedge on auto mode is the most useful sentence in the whole documentation set, and I’d rather quote it than summarise it: “no defense is perfect and no mode replaces your judgment.” They go on to say that for work with real consequences you should stay close and review what Claude does, or switch back to manual approval.
If you wouldn’t leave it unattended with a competent new starter, don’t leave it unattended with an agent.
Four more limits worth carrying into any internal conversation about this.
On what to watch while it works, Anthropic’s advice is better than the obvious version. Rather than trying to validate every command, watch for pattern breaks: files or websites you didn’t mention, or scope creeping past what you asked for. That’s a realistic instruction. Approving every step is a thing people do for two days and then stop.
On data, the distinction is worth being precise about with your IT team. For Team and Enterprise, Anthropic states conversation data is handled under the same commercial commitments as other Team and Enterprise data and “isn’t used to train Claude.”[10] Consumer plans work differently and depend on a setting, so if that question matters in your organisation, the plan you’re on is the answer, not the product.
Everything above, compressed into the thing you’d actually work through. I’ve filled it in with a real example rather than leaving the boxes empty, because a blank template teaches you nothing about the standard.
| Question | Worked example |
|---|---|
| What does finished look like? | “One Word document, same structure as last month’s, with the four channel tables filled in and a short paragraph under each explaining what moved.” |
| How will I know if it’s wrong? | “I have last month’s version and the raw exports. I can check any number in under a minute.” |
| Which folder, exactly? | A new folder, claude-workspace/monthly-report, containing copies of six exports and last month’s report. Nothing else. |
| Which connectors, and why? | None for the first run. Everything it needs is in the folder. |
| Which mode? | Manual, for the first two runs. Auto after that, once I know where it asks. |
| What must it never do? | Send anything. Write into the shared drive. Touch any file outside that folder. |
| Who reviews the output? | Me, before it goes anywhere, every time, for the first month. |
| When do I decide whether to keep it? | After four runs. If it still takes me longer to check than to write, it doesn’t stay. |
A filled-in scoping template for a first Claude Cowork task. The two rows people leave blank are “what must it never do” and “who reviews the output”, and those are the two that decide whether this becomes a habit or an anecdote.
Three things to check before you start that aren’t about the task at all:
Then pick the most boring recurring thing on your desk. The report nobody enjoys, the folder nobody has renamed, the export you reshape by hand every month. Give it a folder of its own, run it in manual mode, and grade the output against the version you’d have produced yourself.
Four runs is usually enough to tell whether the workflow is earning its place. If you’re still checking every line at run four, that’s a real answer and it’s worth having. If you’re skimming by run three because it keeps being right, you’ve found the first job to hand over, and you’ll have a much better instinct for the second.
Claude Cowork is a mode inside the Claude apps, sitting next to Chat, where you hand over a whole task rather than asking a question. Anthropic describes it as using the same agentic architecture that powers Claude Code without needing a terminal, so that instead of responding to prompts one at a time, Claude takes on complex multi-step tasks and executes them on your behalf. Practically, the difference is who does the carrying. In chat you get an answer and then copy, paste and assemble the result yourself. In Cowork it works across the folders and tools you’ve connected, runs code in an isolated environment, and hands back completed files for review. Note that Microsoft ships a separate product called Copilot Cowork, so use the full names.
Multi-step work that’s file-heavy and where the output is a document or a set of files rather than a message. Turning a folder of messy exports into one clean summary. Renaming and sorting hundreds of assets by a rule. Producing a recurring report from source data and last month’s version. Pulling material from a connected app and assembling it into a completed result. The strongest predictor of a good first task is not difficulty, it’s whether you can grade the output quickly. Anthropic also notes that agentic tasks consume more of your usage allocation than chatting, so simple questions are better left in the chat window.
Anthropic’s own guidance is to avoid granting access to local files with sensitive information, and to create a dedicated working folder rather than granting broad access. That’s the vendor’s position, not a critic’s. The architecture does isolate code execution in a sandbox, but Anthropic pairs that with a clear caveat: isolation limits where Claude’s code runs, not what Claude reads or does. Within an attached folder, the agent can reach every file the operating system account can reach. On training data, Anthropic states that Team and Enterprise conversation data isn’t used to train Claude under the same commercial commitments as other Team and Enterprise data, while consumer plans depend on a setting. So the plan you’re on answers part of the question and your folder discipline answers the rest.
No. The whole design point is that it brings agentic capability to knowledge work without a terminal, and it handles ordinary business file types including Word documents, PDFs, spreadsheets, presentations, images and plain text. It does write and run code behind the scenes to do things like reshaping data or generating charts, but you never see or write any of it. What the tool asks for instead is a different skill, which is scoping. Describing what finished looks like, saying which folder it can work in, deciding what it must never do, and knowing how you’ll check the result. That’s closer to briefing a contractor than to programming.
The names are genuinely confusing, and it goes a step further than coincidence: Microsoft’s Copilot Cowork, in general release since June 2026, by default runs on Anthropic’s own Claude models behind the scenes, alongside OpenAI’s models as an option and a Microsoft-trained model on the way. So the two products can share an engine without being the same product. The practical differences are where your files live, what the work is grounded in, and how you pay. Claude Cowork reaches folders on your own machine through the desktop app, plus an open directory of connectors, included in your Claude subscription. Copilot Cowork has no local folder access at all: it runs in the cloud, grounded in whatever your organisation already holds in SharePoint, Outlook and Teams, and it’s billed on top of a Microsoft 365 Copilot licence using usage-based credits rather than a flat fee. Claude Cowork is more file- and connector-oriented across different services. Copilot Cowork is deeply Microsoft 365-native and tenant-centred.
Every claim about Claude Cowork here was checked against Anthropic’s own current support and product documentation on 25 August 2026 rather than against press coverage, because much of the coverage still circulating describes a January 2026 research preview that was Mac-only and limited to the top plan, which stopped being accurate in April 2026. Several things are deliberately not stated because they could not be confirmed. Anthropic publishes no numeric usage limit for Cowork, only that it consumes more allocation than chat, so no figure appears here. Its product page and its support documentation currently contradict each other on whether Cowork activity is captured in the Compliance API, so this article takes no position on that. The connectors gallery has no Cowork filter, so no specific connector is described as Cowork-compatible. This piece deliberately doesn’t quote specific dollar figures, because Anthropic’s own pricing page carries a note that prices are subject to change and a number printed here would go stale faster than the article gets updated. The Microsoft Copilot Cowork comparison, including the detail that it runs on Anthropic’s Claude models by default, is drawn from Microsoft’s own product blog and admin documentation rather than from comparison articles.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.