Claude Cowork

Claude Cowork is a program made by the American company Anthropic. Anthropic makes Claude, a computer program that understands and writes text and responds like a conversation partner. Most such programs work like this: you write a question, get an answer, then write the next question. Cowork goes a step further. You describe an entire task, and the program works through it itself in many individual steps. In doing so, it opens files on the computer, changes them, and creates new ones. The human watches and steps in if something goes wrong.

From answer-giver to co-worker

The difference sounds small, but it matters a lot in practice. A chat window delivers text that you then have to copy, file away, and process yourself. These manual steps often cost more time than the thinking itself. Cowork takes over exactly this tedious part.

With this, Anthropic is targeting people who cannot program. So far, it has mainly been software developers who benefited from such tools, because their work already consists of text and files anyway. Cowork is meant to do the same for office work: checking spreadsheets, structuring reports, searching a hundred documents for a keyword. This is why experts are watching the product closely.

Economically, there’s a simple calculation behind this. The longer such a program works on a task, the more computing power it consumes. Providers like Anthropic earn money from this computing time. Tools that work independently for hours are therefore also more commercially attractive than short question-and-answer exchanges.

What happens when a task is worked through

Cowork is given access to a folder with files. From the task, it first derives a plan with several sub-steps. It then works through these steps one after another and checks after each step whether the result fits. If something is wrong, it corrects itself and tries a different approach.

This is made possible by so-called tools. These are small additional functions that the program is allowed to call: read a file, write a file, search the web, run a calculation. The language model itself can only generate text. It is only these tools that turn text into an actual action on the computer.

An important point is control. For sensitive actions, the program asks for confirmation, for example before deleting files or sending something out. However, one should not rely on this blindly. A well-known risk is called prompt injection: a document contains hidden instructions that the program mistakes for a task. This is why the rule applies to only give such tools access to data whose loss you could live with.

Cowork compared to other assistants

At the start of 2026, Cowork is a young application and initially available only to paying Claude subscribers. Anyone trying it out works not in a chat window but in an interface with projects and files. Comparable approaches also exist from other providers, such as agent features from OpenAI or Google.

In news coverage, the name usually comes up in connection with the term agent. This is what AI programs are called that independently plan and execute multiple steps. Cowork is considered an example of this technology moving beyond the field of programming.

A common misconception is that Cowork is a new, smarter language model. That is not true. It uses the same Claude models as the regular chat. What’s new is only the packaging: access to files, a plan, and permission to keep working alone for a long time.