What is Claude Cowork, and how does it let Claude work in your files?

Yesterday, Monday, January 12, 2026, Anthropic released Cowork, a new way to use Claude in its desktop app for Mac. In an ordinary chat, you ask and Claude answers in text. In Cowork, you point Claude at a folder on your computer and give it a job, and it works through the steps itself: opening files, sorting them, writing new documents and telling you what it is doing along the way.

Anthropic's own short description is "Claude Code for the rest of your work." Claude Code is the company's coding agent, a tool that works inside a programmer's project from the command line. Anthropic noticed that many people were using it for tasks that had nothing to do with code. Cowork offers the same kind of help without the technical setup.

What Anthropic released

Cowork is a research preview, an early version released to learn how people use it. Here are the details:

Anthropic's examples are everyday office chores: tidying a downloads folder by sorting and renaming each file, turning a pile of receipt screenshots into an expense spreadsheet, or drafting a report from scattered notes.

How Cowork works

An agent, in AI, is a system that pursues a goal by planning and taking actions over several steps, rather than replying once. Cowork is Claude acting as an agent on your own files.

When you give it a task, Claude first makes a plan, then carries it out step by step while keeping you informed. TechCrunch reports that the Claude Agent SDK sits underneath Cowork. That is Anthropic's toolkit for building agents, so Cowork relies on the same underlying model and machinery as Claude Code.

Three things extend what it can do:

• Connectors link Claude to outside services and information, such as your other apps.
• Skills are packaged instructions that help Claude do a specific kind of work well. Cowork arrives with a first set for creating documents, presentations and other files.
• Claude in Chrome, Anthropic's browser extension, lets Cowork complete tasks that need a website.

You also don't have to wait. You can queue up several tasks, which Claude works through in parallel, or offer more ideas while it is still busy. Anthropic compares the experience to leaving messages for a coworker rather than holding a back-and-forth conversation.

The folder is the key safety idea. Claude can see and change only what you explicitly share. TechCrunch describes this as a sandboxed version of Claude Code: a sandbox is a fenced-off area where software can act without touching the rest of your system.

The risks, stated plainly

Anthropic is direct about what can go wrong. Claude asks before taking significant actions, but it can take destructive actions, such as deleting files, if your instructions tell it to, and it might misread an unclear request. The company's advice is to give very clear guidance, especially about anything that can't be undone.

The second risk is prompt injection: hidden instructions placed in content Claude reads, for example on a web page, that try to change its plan. Anthropic says it has built defenses, but that securing an agent's actions in the real world is still an unsolved problem across the industry. As the announcement puts it, "These risks aren't new with Cowork, but it might be the first time you're using a more advanced tool."

A worked example: receipts to a spreadsheet

Suppose you have a folder called "Trip receipts" with photos and PDFs from a work trip. You open Cowork, share only that folder and write:

In this folder are receipts from my trip to Toronto, March 3 to 6.
1. Read every receipt and make a spreadsheet called expenses.xlsx with
   date, vendor, category (travel, meals, lodging, other), amount and
   currency.
2. Add a total per category and a grand total.
3. List any receipt you could not read in a separate sheet.
4. Do not delete, rename or move the original files.

Claude would plan the job, open each receipt, extract the details and build the spreadsheet, telling you as it goes. The last line sets a clear boundary, and the third catches anything it couldn't handle. Before submitting the report, you would still check a few amounts against the originals.

Why it matters

Cowork moves an AI agent from the programmer's terminal to the ordinary desktop. Without learning any commands, a non-programmer can hand Claude a folder and a goal.

It also reflects a pattern in how people use AI. Anthropic built Cowork because users had already stretched Claude Code to tasks it wasn't designed for, from managing media files to analyzing conversations. Building a simpler front door for that behavior is a bet that much everyday office work is a series of file tasks an agent can do.

Anthropic is releasing it early on purpose, to see how people put it to work, and says it will improve it quickly and look for further ways to make it safer.

What is a good first task to give Cowork?

Choose a task that is useful, easy to check and harmless if something goes wrong. Organizing a copy of a folder, summarizing a set of notes into a single document, or building a spreadsheet from receipts are all good starts. Avoid anything involving deletion, money or sensitive personal records until you know how it behaves.

Share the smallest folder that will do the job, and work on a copy when you can. In your instructions, state the goal, the format you want and what Claude must not touch. Watch its progress updates during the first runs, and answer its questions carefully when it asks before a significant step.

When it finishes, review the results as you would a new assistant's work. Open the files it created, check a sample of the details, and only then move on to longer or more important jobs.