What Is Claude Cowork? How It Actually Works

Claude Cowork is Claude working directly inside your files, folders, and apps instead of just chatting about them — it plans a task, shows you the plan, then reads, edits, and produces real files on your machine while you stay in control of what it's allowed to touch. This guide covers exactly how it works: folder scoping, plan approval, skills, scheduled tasks, and connectors, plus a first-automation walkthrough for people who have never opened it.

Everything below is written for non-technical professionals — operators, marketers, consultants, founders — not developers. If you're deciding whether to pay for it, our Claude Cowork pricing guide breaks down which plans include it. If you want ready-made automation ideas once you're set up, see Claude Cowork automations.

What is Claude Cowork?

Claude Cowork is an agentic workspace, built into the Claude desktop app, that lets Claude act on a folder of real files and apps rather than only answering questions in a chat window. You describe a job in plain English — "reconcile these three spreadsheets and flag anything that doesn't match" — and Cowork breaks it into a visible plan, works through the steps, and leaves finished files behind instead of a wall of text to copy out yourself.

The distinction from regular Claude chat is the difference between describing work and doing it. Chat is a conversation; Cowork is a working session. Anthropic launched it in February 2026 as a desktop-only tool and has since expanded it to web and mobile, so a task you start on your laptop can keep running in the background and you can check on it from your phone. Internal usage data Anthropic has shared shows the biggest category of Cowork tasks is business process work — not code — which is exactly the non-technical audience this guide is written for.

If you want the fuller side-by-side on when to reach for which one, we cover it in Claude Cowork vs Claude Code.

Folder scoping: Cowork's core safety boundary

Folder scoping means Claude never gets blanket access to your computer — you point it at one specific folder, and everything it can read, write, or run stays inside that boundary. Instead of turning Claude loose on your entire hard drive, you hand it something like ~/Documents/q2-launch/, and from that moment on, its reach stops at the folder's edge.

This matters because it is the mechanism that makes the rest of Cowork trustworthy. Once a folder is scoped, Claude can read every file inside it, create new ones, edit existing ones, and run scheduled tasks against that content — but it cannot wander into your Downloads folder, your other client's workspace, or system files it was never given. The practical habit worth building: create one folder per project or client, the same way you'd create a separate Claude Project for each engagement, and only ever scope Cowork to the one you're actively working in.

Scoping reduces lateral exposure, but it does not remove every risk inside the folder itself — Claude can still overwrite or delete a file it was scoped to touch. That is exactly what plan approval is for.

Plan approval: the three modes that decide what Claude can do

Before Cowork touches anything, it turns your instruction into an explicit, readable plan — the list of steps it intends to take — and you decide how much freedom it gets to run that plan without stopping to ask. That control lives in a mode selector in the chat box, and it applies to actions like using a connector, not just to editing files.

Even inside an approved plan, Cowork will still stop and check with you before certain sensitive actions — downloading a file or entering information into a page it hasn't seen before are common examples. It also will not silently deviate from the plan you approved; if the task changes shape mid-way, it asks again before continuing.

Skills and plugins: how Cowork learns your workflows

A skill is a folder containing a SKILL.md file of instructions that Claude loads automatically whenever it's relevant to the task at hand — write it once and it works the same way across Claude chat, Cowork, and Claude Code. Instead of re-explaining your invoicing format or your report structure every single time, you write it down once as a skill and Cowork picks it up on its own.

Plugins are the bigger unit: a plugin bundles multiple skills, connectors, and even sub-agents into one package built for a specific job — financial analysis, engineering review, HR onboarding. Anthropic's own teams use domain-specific plugins internally to encode institutional workflows so a new Cowork session starts already knowing how that part of the business operates, rather than starting from a blank instruction every time.

A minimal skill file for a recurring reporting task looks like this:

# SKILL.md
name: weekly-status-report
description: Use when compiling the weekly team status update.

When asked to build the weekly status report:
1. Read every file in /updates/ dated within the last 7 days.
2. Group items by project, not by person.
3. Flag anything marked "blocked" at the top of the report.
4. Write the output to /reports/weekly-status-[date].md
5. Keep it under 400 words. No filler, no restating the obvious.

Drop that file in the scoped folder's skills directory and every future "build the weekly status report" request in that folder follows the same structure automatically.

Scheduled tasks: automation that keeps running after you close the laptop

A scheduled task is a Cowork job set to a human-readable cadence — "every weekday at 8am", "first Monday of the month" — that runs in the background and writes its results to disk without you needing to be present. With Cowork's expansion to web and mobile, tasks can now run with no device online at all: you schedule the work, close everything, and come back to a finished file.

This is the feature that turns Cowork from "a faster way to do a task" into "a task that stops being your job." A weekly competitor price check, a Monday-morning inbox triage, a month-end expense reconciliation — anything with a fixed rhythm is a candidate. We go deep on specific setups worth automating first in Claude Cowork automations.

Connectors: bringing your existing tools into Cowork

Connectors are how Cowork reaches outside the local folder into the apps you already run your business on — Google Drive, Gmail, DocuSign, and similar services — and you control exactly which ones are switched on through the + menu in the chat box or the Customize > Connectors page. Nothing connects by default; each one is an explicit opt-in, scoped the same deliberate way folders are.

On the organization side, Anthropic has also opened Cowork up to enterprise systems through a plugin framework — teams can package a set of connectors and skills together so an entire department starts from the same vetted configuration instead of everyone wiring up their own connections by hand. For a solo user or small team, the practical version of that same idea is simpler: decide once which two or three tools a given folder actually needs, connect only those, and leave the rest off.

Combined with plan approval, connectors are where the safety model does its most important work: a connector gives Cowork reach into a live system, and the approval mode governs how much it can do there without checking in first. For a first project, it is worth turning on exactly one connector — whichever tool the task actually needs — rather than switching everything on at once. Once you have watched Cowork use that one connector safely a few times, adding a second one is a much smaller leap of faith than starting with all of them enabled.

Who Claude Cowork is actually for

Cowork is built for the person who has a folder of real files and a recurring, describable job to do with them — not for people writing software. Anthropic's own usage breakdown backs this up: the largest share of Cowork tasks is business process operations, with content creation and copywriting close behind, and software development a distant third. If your week involves spreadsheets, reports, client folders, or recurring admin, Cowork was built with you in mind. Our broader Claude for business guide covers where Cowork fits alongside Claude's other business tools.

In practice, that means a good Cowork candidate looks like this: there is a folder that already exists (client files, weekly reports, research notes), the task is something you have done more than once by hand, and the output is a file, not a conversation. Operations managers reconciling data across spreadsheets, marketers repurposing a folder of assets into channel-specific formats, and consultants who keep a working folder per engagement all fit that pattern directly. A one-time question, by contrast, is still better handled in plain chat — spinning up a scoped folder and a plan for something you will never repeat is more setup than the task deserves.

It is a poor fit for a single quick question — that is what chat is for — and it is not a code editor, even though it can technically touch code files. If your work is genuinely software development, our Cowork vs Claude Code comparison will point you to the right tool faster than trial and error will.

Your first Cowork automation, step by step

The fastest way to understand Cowork is to run one real task end-to-end rather than reading about it further. Here is the setup that gets a first-timer to a finished, useful result in under 20 minutes.

A prompt that works well for that first run, adjusted to your folder:

Look at every file in this folder. Build a single summary document
that lists:
- what each file is and its last-modified date
- anything that looks incomplete, outdated, or inconsistent with
  the others
- a short recommended next action for each flagged item

Save the result as summary-[today's date].md in this folder.
Show me your plan before you start.

Frequently asked questions

Does Claude Cowork cost extra on top of Claude Pro?

Availability depends on your plan, and Anthropic has been rolling Cowork out gradually across desktop, web, and mobile. Our Cowork pricing guidehas the current breakdown by plan so you're not guessing.

Can Cowork delete or overwrite files I care about?

Yes, within whatever folder you've scoped it to — that is exactly why plan approval and starting in Manually approve mode matter. Folder scoping limits where Claude can act; it does not remove the need to review what it's about to do inside that folder, especially the first few times.

Is Claude Cowork the same as Claude Code?

No. Claude Code is built for software development in a code repository; Cowork is built for non-technical, file-and-folder work like reports, research, and admin. See Cowork vs Claude Code for the full comparison.

What happens if a scheduled task goes wrong while I'm away?

Cowork writes results to disk rather than taking irreversible external actions unless you've explicitly approved that connector and mode combination. Start any new scheduled task in Manually approve or Automatically approve mode until you've seen it run correctly a few times, then loosen the mode once you trust it.