Cowork can work in the cloud. Your folder is still on the laptop.

Over the summer Cowork gained the ability to run tasks in the cloud. You can ask it to go through the inbox and write a meeting brief, close the laptop and come back to the result later. The job can carry on without your machine being on, as long as it does not need local files or tools.

That is the practical difference with Cowork in the cloud. You can start a task from the phone and come back to the result later. But a setup that uses a folder on your computer does not become independent of the computer just because the agent runs somewhere else.

In An agent in a folder we wrote about gathering sources, rules, examples and deliverables in one bounded workspace. That way of working is still relevant. Now we also have to separate where the agent runs from where the material it needs lives.

What has moved to the cloud?

In a cloud session in Cowork the work and the code execution run on Anthropic's servers. That holds when you start the cloud session from the desktop app too. The app you are using does not on its own tell you where the job is running.

The session follows your Claude account. You can start on the laptop, follow up from the phone and collect the result on the web. For tasks that only use connected services and files saved to the account, the laptop does not need to be on.

This is honestly rather confusing for many people right now. It is not always easy to tell where the job runs, which files Claude can reach, and what is kept for next time.

The folder does not move on its own

A folder on your machine does not become available in the cloud by itself. Cowork can reach it through the desktop app, but only within the folders and permissions you have granted.

Anthropic is specific that a cloud session has to be started from the desktop app if it is to read and write local files. The app has to stay open and connected on that machine. After that you can follow the session from the web or the phone.

Close the app and the cloud session carries on, but it loses access to the local files. The same dependency applies to local integrations and to steering the browser or other programs on the machine. An ordinary web search is one thing. Letting Claude click around in a browser on your machine is another.

It also means that folder work is not necessarily the same as local execution. The agent can run in the cloud and use the desktop app as its link to your files. The machine is still part of the setup.

Projects and folders

Chat and Cowork now share the same entrance, but it still matters to separate what lives in the cloud from what lives on your machine.

In the cloud, projects and delivered files are saved to the Claude account. Think of it as a simple place for files and results, a little like OneDrive or Google Drive, though not a full replacement. A project gathers instructions and reference material you can use in both chat and Cowork, across devices.

On your machine sit local folders with files Cowork can read and edit through the desktop app. If you create a project from a local folder, it stays on the machine and is used in Cowork on the desktop. It does not move to the cloud on its own.

In short, the account project follows you between devices and the local folder does not. And if you want to keep a finished report as knowledge in a cloud project, you add it yourself.

What is kept in the cloud?

When Cowork works in the cloud it uses a temporary workspace on Anthropic's servers. Think of it as a workbench rather than a file archive. The session and the files Claude delivers are saved to the account, but working files that are not delivered are not kept after the session.

Ask Claude to deliver whatever you want to use again, scripts included. If a report is to become standing reference material in a cloud project, you add it yourself. Download or save important deliverables in the company's usual file store, so you do not depend on them staying in Cowork.

Scheduled cloud tasks need sources they can reach without the laptop

Scheduled tasks in the cloud can run even when the computer is off. Each run becomes its own Cowork session, and you read the result when it suits you.

But these tasks cannot use local folders or local integrations. They have to fetch the material from connected services or from files saved to the Claude account.

Here is an important exception. Existing scheduled tasks that depend on local files or integrations can still run locally. They do not become cloud tasks on their own, and they still depend on the machine.

A weekly report that reads a folder of Excel exports on the laptop cannot simply be moved to a schedule in the cloud. You either make the updated exports available through an approved cloud service or keep the local run.

This builds on Connectors and automation in Cowork. The integrations give access to the sources. The schedule decides when the job happens. For it to work with the laptop closed, every source and tool it needs has to be reachable without it.

Microsoft 365 got an important update too

The timing is good for companies on Microsoft 365. On 7 July the connector gained support for making changes as well as searching and reading. With the write tools enabled Claude can, among other things

• draft, send and organise email
• create and change calendar appointments
• create and update files in OneDrive and SharePoint

Together with Cowork in the cloud this opens up work that does not need a laptop switched on. A meeting brief can build on Outlook threads and documents in SharePoint, and a weekly report can be saved back to OneDrive for review. Both the sources and the deliverables can live where the company already works.

An administrator has to approve the new permissions and enable the write tools. Start with reading and drafts, and keep control of sending and other actions that have consequences.

Two kinds of tasks to keep apart

I would use this as a simple split when you weigh what fits in the cloud.

1. Tasks that do not need your machine. A morning brief from email and calendar, an overview of open customer threads, or a report built on documents in a connected service. Here the cloud is a natural place to start, provided the integrations give the access the task needs.
2. Tasks that need something on your machine. A proposal built on local templates and past offers, editing files in a working folder, or a job that needs a local program. Here the desktop app still has to be available, even when the session itself runs in the cloud.

This is not a split between simple and advanced tasks. A cloud session can analyse data and produce documents too. The question is whether it can reach everything it needs.

If you have used Dispatch from the phone, it is worth keeping the two apart as well. Dispatch sends work to your computer and needs it awake with the desktop app open. A cloud session without local dependencies does not.

Memory helps, but does not replace the method

With the new memory feature, chat and Cowork in the cloud share what Claude has stored as memory. That might be how you like a summary written. It requires memory to be switched on, and it does not mean every detail from every conversation follows along.

Projects have their own memories. On Team and Enterprise, memory is off by default and has to be enabled by an owner before each user can turn it on. Anthropic has moved users off the older memory experience, and memory is not available to organisations with agreements that limit it, such as HIPAA, public-sector or custom data retention terms. Local Cowork sessions do not use this shared memory.

I would still write down the method for tasks that are going to repeat. What should the report contain? Which sources come first? What always has to be checked before something is done?

This is where skills in Cowork are useful. They gather instructions, examples and any scripts into a way of working you can use again. Skills are available in the cloud too, but a skill that depends on a local tool does not become independent of the laptop for that reason.

Check access, data and actions

Cowork in the cloud is in beta. Web and mobile are available on Pro, Max and Team, and on Enterprise where an administrator has enabled it. Check what is available and allowed on your account before you build a workflow around it.

And keep reading apart from editing. Writing a summary is a different thing from sending email, changing the calendar or deleting files.

Start with a meeting brief

A concrete first task can be to pull today's meetings from the calendar and the relevant background from email. Ask for a short note, not for Claude to contact anyone on your behalf.

Write a short brief for today's meetings based on the calendar and the relevant email threads. Include links to the sources, open questions and what I should prepare. Mark what you are unsure about. Do not send messages or change the calendar.

Run the task by hand first. Check that the right meeting, the right person and the right thread are linked together. Limit the tool access and use manual approval for actions where a mistake has consequences. A sentence in the task description is no substitute for access control.

When the quality is steady, set up a fixed run with only the access the task needs. Test that it works without the desktop app too, so you find local dependencies before you start relying on the routine.

When you give feedback, save the improvements in the instructions or in a skill. That is the link to what we wrote about building AI capital. The value lies in the method getting better each time, as much as in the task running while you do something else.

Other agents worth testing

Cowork was not alone in getting a big update over the summer. OpenAI launched ChatGPT Work, an alternative to Cowork that can work with tools and files and deliver documents, spreadsheets and presentations.

We can warmly recommend both, even though we have a slightly special relationship with Claude. Cowork was where we first saw agents become properly good for most people.

Anthropic and OpenAI take turns at the front, and right now what each is best at varies. GPT-6 Astra is a remarkable model, and our first round of testing put it clearly ahead of Claude in some areas. You really have to test for yourself and see what works best for your tasks.

Microsoft made Copilot Cowork generally available to Microsoft 365 Copilot customers over the summer. Microsoft says it runs on Anthropic models. We find working with files in the cloud smoother there, especially with SharePoint, but the agent experience itself is not quite at the level of Claude Cowork and ChatGPT Work yet.

What is your experience with agents in the cloud? What do you use them for, and what works best?