What Cowork is, and how it differs from a chat

In a chat you ask a question and read an answer. Whatever happens next is your job. Cowork inverts that: you describe an outcome, and Claude works towards it across your files and connected tools, showing you each step, and hands back finished work for review.

That is a different kind of thing, and it deserves a different kind of caution. A chat that is wrong wastes your time. An agent that is wrong has already edited the file.

What it actually does

It works in folders you connect, reaches into live systems through connectors, drives a built-in browser, and can be given your screen as a last resort. The current feature list lives in Anthropic's getting-started article and moves faster than this page.

The property worth carrying is the order. Claude reaches for the most precise tool first: a connector where one exists, then the browser, and only then screen interaction. Each rung outward is a wider grant, which is why the three grants have their own page.

Where it runs

Two modes, and the difference matters more than it looks.

Cloud sessions are the default. The agent loop and code execution run on Anthropic's servers, in an isolated sandbox created when the session starts and destroyed when it ends. Sessions and files are saved to your Claude account.

Local sessions run the agent loop on your own device, with code execution inside an isolated virtual machine — Apple's Virtualization.framework on macOS, Hyper-V on Windows.

Either way, when Claude needs something on your machine the request goes through the Claude Desktop app over an Anthropic-brokered connection, and each local tool call is checked against your permissions before it runs.

Where you can use it

Paid plans, across desktop, web, mobile and a Chrome side panel. Which plan gets which surface is a matrix that moves, and the product page carries the current one. What the phone is good for, and what stays tethered to the desktop, is its own page: steering from away.

What goes wrong

Treating it as a faster chat. The instinct is to give it a small task to try it out, which is fine, and then to keep giving it small tasks, which wastes it. Cowork earns its cost on multi-step work: research synthesis, document production, pulling data from several places into one artefact. A question with a one-paragraph answer belongs in a chat.

Not noticing the usage cost. Cowork consumes considerably more of your allowance than standard chat, because it is doing considerably more. If you run it casually all afternoon you will meet a limit you have never met before.

Assuming "it showed me the steps" means "I reviewed it." Watching a progress indicator is not review. See Reviewing agent output you did not watch being made.

How to check it worked

Before trusting Cowork with anything that matters, run it once on a copy of a folder you would not mind losing, and read the finished result against the source material yourself. What you are checking is not whether it can do the task — it can — but whether you can tell from the output when it has gone wrong.