Plain-English Explainer
What Is Claude Cowork? The Folder and Permission Model Explained
Not a chat window that talks about your work, but a coworker that opens the folder and does it.
Claude Cowork is a mode inside the Claude desktop app that takes on multi-step tasks using your own files and connected apps. This page explains what it is, what it needs, how the folder and permission model keeps you in control, where its limits are, and who gets the most from it.
Claude Cowork is a mode in the Claude desktop app, available on macOS and Windows, that works on tasks using your own files and connected apps. It requires a paid plan (Pro, Max, Team, or Enterprise). Claude sees no files until you connect a folder, and three permission modes (Manual, Auto, and Skip) control how often it pauses for approval. Permanently deleting files always requires your explicit permission. Instead of only answering in a chat, Cowork reads and writes files in connected folders directly and carries a task through multiple steps.
Published
Free lesson · no account neededMeet ClaudeWhy Claude?6 minStart the free lessonKey Takeaways
- Cowork is a mode in the Claude desktop app for macOS or Windows, not a separate product. It works on tasks using your own files and connected apps rather than only replying with text.
- It needs a paid plan: Pro, Max, Team, or Enterprise. On Team or Enterprise, an admin may need to enable Cowork for the organization before you can use it.
- Claude sees no files until you explicitly connect a folder. You are prompted the first time you ask it to work with files, and you can connect more folders later.
- Manual, Auto, and Skip modes set how often Claude pauses for approval. Deleting files permanently always needs your explicit Allow, whichever mode you pick.
- A good first task is something you would otherwise do yourself, such as a one-page brief from a folder of PDFs or a cleaned and charted spreadsheet.
Cowork's three permission modes side by side
| Mode | How it behaves | Default in the newer experience | When it fits |
|---|---|---|---|
| Manual | Claude pauses and asks for your approval before each action. | Default on Enterprise. | New tasks, unfamiliar folders, and work where you want to see every step before it happens. |
| Auto | Claude keeps working without asking at every step. A safety review checks each action for risks like data exfiltration or prompt injection and blocks unsafe ones. | Default on Pro, Max, and Team. | Routine tasks in folders and connectors you know, when you want momentum without constant approvals. |
| Skip | Nothing pauses and nothing is checked automatically. | Not a default. | Only when you completely trust every action, connector, file, and app involved, which is the guidance in Anthropic's safety article. |
Copy and Paste
Prompts You Can Use Today
Turn a folder of PDFs into a one-page brief
I have connected the folder Q4-vendor-proposals. Read every PDF in it and write a one-page brief for my director, who has five minutes and no background. Cover what each vendor offers, the price they quote, the biggest risk in each proposal, and your recommendation. Save the result as a Word document in the same folder and list any PDF you could not read.
Swap the folder name, the reader, and the sections for your own situation. Keep the instruction to list files it could not read.
Draft a status update from a Slack channel
Using the Slack connector, read the last week of messages in the project-atlas channel. Draft a status update for stakeholders who do not follow the channel. Use three headings: Done this week, In progress, and Blocked or at risk. Keep it under 250 words, name owners where the messages make them clear, and flag anything you are unsure about instead of guessing.
Replace the channel name and the time window. If you would rather review before anything is shared, tell Claude to produce a draft only.
Clean a messy spreadsheet and chart the results
Open the file signups-export.xlsx in my connected folder. Standardize the date column, remove exact duplicate rows, fix inconsistent capitalization in the country column, and tell me every change you made. Then build a chart of signups per month and save the cleaned file as a new copy so the original stays untouched.
The line about saving a new copy protects your original. Rename the columns to match your own file.
Draft your global instructions
I want to write global instructions that apply to every Cowork session. I am a marketing operations manager who writes for busy executives. Interview me with five short questions about my tone, preferred output formats, and things I never want you to do, then turn my answers into a block under 200 words I can paste into Settings.
Replace the role with yours. The interview step gets better instructions than guessing at them cold.
Set up folder instructions for a recurring project
This connected folder holds my monthly board-report materials. Look at the files already here and propose folder instructions covering how files are named, which template to start from, and what the report structure usually looks like. Show me the draft first and wait for my edits before saving it as the folder's instructions.
Replace the project description. Claude can update folder instructions itself over time, so review them now and then.
Ask for a plan before any files change
I want to reorganize the Downloads folder I connected into subfolders by year and project. Before you move or rename anything, give me a table of the proposed moves with the current path and new path. Do not delete anything. Wait for my approval of the table, then carry it out in batches and report what changed.
This pattern works in any permission mode and is a good habit for any task that touches many files.
The core ideas
What Claude Cowork does differently from a chat window
A normal chat gives you text back and leaves the work to you. Cowork is built so that Claude does the work in the places where your work already lives.
Works on your own files
On desktop, Claude reads and writes local files in the folders you connect, directly. There is no uploading a document and downloading the result.
Reaches your connected apps
MCP connectors such as Slack, Gmail, Google Drive, Asana, and GitHub let a task pull from the tools you already use. Claude may suggest a connector when a task calls for one.
Organizes work into projects
A Cowork project is a workspace tied to a specific folder, with its own instructions, files, and memory, so recurring work does not start from zero.
Lets you watch or step away
Claude explains each step and its reasoning. You can watch, intervene mid-task to redirect it, or leave and come back to finished work.
Keeps you in control
Folder access is opt-in, permission modes set how often Claude pauses, and permanent deletion always needs your explicit permission.
The definition
Claude Cowork in plain terms
Claude Cowork is a mode in the Claude desktop app. It is not a different model and not a separate download from the app itself. It is a way of using Claude in which you describe an outcome, and Claude works toward it using your own files and connected apps, instead of only replying inside a conversation.
The easiest comparison is a capable colleague who sits down at your computer, but only in the folders you point them to. You say, in ordinary language, that you need a brief from the proposals in a folder, or a status update from a Slack channel, or a messy spreadsheet cleaned and charted. Claude then plans the steps, does them, and shows you its reasoning along the way.
That is the practical difference from asking the same question in an ordinary chat. In a chat, you paste text in and copy text out. In Cowork, the inputs are files and apps that already exist, and the outputs are files that land back in your connected folder, ready to open in whatever program you normally use.
Cowork runs on the Claude desktop app for macOS or Windows. If you have only used Claude in a browser tab, the desktop app is the first thing to install. Because Cowork needs access to local folders, the desktop app is what makes the folder model described below possible.
Requirements
What you need before you can use it
Cowork requires two things: the Claude desktop app and a paid plan. The plans that include it are Pro, Max, Team, and Enterprise. If you are on a free plan, the mode is not available to you until you move to one of those.
On Team or Enterprise plans there is one more gate. An admin may need to enable Claude Cowork for the organization before a member can use it. If you are a member and the mode is missing, the likely first question is not a technical one. It is whether your organization has turned it on, so ask whoever manages your Claude workspace.
Team and Enterprise owners also hold some organization-level controls. For example, they can disable web search for Cowork, Chat, and Claude in Chrome in organization settings. That matters if your company wants Claude to work on internal files without reaching out to the open web.
The folder model
Claude sees nothing until you connect a folder
The most important sentence about Cowork's access is also the simplest: Claude sees no files until you explicitly connect a folder. Installing the app does not expose your Documents directory. Nothing is read from your disk until you grant access to a specific place.
The first time you ask Cowork to work with files, you are prompted to connect a folder. You can connect more folders later as your tasks need them. This makes the sensible habit obvious: connect the folder that holds the task's materials, not your whole home directory. A folder named for a project is a better grant than a folder that holds everything.
Once a folder is connected, Claude reads and writes there directly. That is what makes the workflow feel different. You do not attach files to a message, wait for an upload, then download a result and drag it back to where it belongs. The brief or cleaned spreadsheet simply appears in the folder.
A practical consequence follows. Because Claude can write into a connected folder, it is worth asking for new copies instead of edits whenever a file matters. Telling Claude to save a cleaned spreadsheet as a new file leaves the original untouched, and that costs you one sentence in the prompt.
The permission model
Manual, Auto, and Skip: choosing how much to approve
Cowork has permission modes that you can change from a mode selector in the chat box. In Manual mode, Claude pauses and asks for approval before each action. In the newer experience it is the default on Enterprise. It is the mode to start with when a folder or task is new to you, because you see every action before it happens.
In Auto mode, Claude keeps working without asking at every step. A safety review checks each action for risks like data exfiltration or prompt injection and blocks the ones it judges unsafe. In the newer experience it is the default on Pro, Max, and Team. It trades some visibility for speed, with a review layer standing in for your approval clicks.
Skip mode is different in kind. Nothing pauses and nothing is checked automatically. Anthropic's safety article says to use it only when you completely trust every action, connector, file, and app involved. If a task reads email, web pages, or files from other people, those inputs are not fully under your control, and that is a reason to stay out of Skip.
Connectors have their own settings. Each connector's permission can be set to Always allow, Needs approval, or Blocked, and a blocked connector tool is denied in every mode. So even if you work in Auto or Skip, you can keep a particular tool off the table. Deletion has a floor as well: Cowork always requires explicit permission before permanently deleting any files, and you must choose Allow at the prompt.
Even in a mode that otherwise does not pause for approval, Claude is guided to seek explicit confirmation for sensitive actions. The examples given are captchas, purchases and financial transactions, sensitive data, and system file changes. Treat that as a backstop, not as permission to stop paying attention.
Safety boundaries
Where Cowork draws its lines
These are the specific boundaries documented by Anthropic, collected in one place.
Isolated execution
Shell commands and generated code run in an isolated environment on Anthropic's servers, separate from your own computer and network.
Deletion needs a yes
Permanently deleting any file always triggers a prompt, and you must choose Allow. No permission mode removes that step.
Blocked means blocked
A connector tool set to Blocked is denied in every mode, including Skip. Use it for tools you never want a task to touch.
Task deletion is real
Deleting a task removes it from your history immediately and from backend storage within 30 days.
Memory is your choice per task
You can turn memory off for a single task before its first message. Once the task has started, that setting cannot be changed.
A task from start to finish
What running a Cowork task looks like
You begin by describing the outcome and naming the materials. A good request says which folder or connector to use, who the output is for, what format you want, and what Claude should flag instead of guessing. The more your prompt reads like a brief to a new colleague, the less you will need to correct afterward.
While the task runs, you can watch Claude work or step away. Claude explains each step and its reasoning, so a quick glance tells you whether it understood the request. If it heads somewhere you did not intend, you can intervene mid-task to redirect it instead of waiting for a finished result you will have to throw out.
For complex work, Claude may run multiple sub-agents in parallel. You do not manage them yourself. The practical effect is that a task with several independent parts, such as reading many documents, can be split up rather than handled strictly one piece at a time.
Long tasks can keep running in the cloud when the desktop app window is closed. There is one catch worth remembering: local file, browser, and computer access require the desktop app to stay open. A task that needs your connected folders is tied to the app being open, even if some work can continue in the cloud.
Making it yours
Projects, global instructions, and folder instructions
Cowork has three layers of memory and guidance. Global instructions, set under Settings > Claude Cowork > Global instructions, apply to every Cowork session. They are the place for things that are always true about you and your preferences, such as tone, output format, or your role.
Folder-specific instructions live in a Customize panel for a connected folder. They are for facts that are true only of that body of work, such as naming conventions, templates, or the usual structure of a report. Claude can update folder instructions itself as it learns how you work, which is convenient, and also a reason to read them occasionally.
A Cowork project ties these together. It is a workspace tied to a specific folder, with its own instructions, files, and memory. For work that repeats, like a monthly report or a weekly update, a project means the second run starts with everything the first run taught Claude.
Memory is something you can decide about per task. If a task involves material you do not want carried forward, turn memory off before sending the first message. After the task starts, the setting cannot be changed, so make the decision at the start.
Honest limits
What Cowork does not do for you
Cowork does not remove the need to check the work. Claude can misread a document, summarize a spreadsheet incorrectly, or draft a status update that sounds confident about something the source did not say. If the output goes to a client, a regulator, or your executives, you are still the person who signs off on it. Asking Claude to cite which file or message each claim came from makes that review much faster.
The egress setting is a good example of a boundary that is narrower than people assume. Cowork respects your configured network egress permissions, but those egress settings do not apply to the web search tool, web fetch, MCP connectors, or Claude in Chrome. If your organization relies on egress rules to control where data can go, read that sentence carefully and use the organization-level controls for web search where they apply.
It also does not work without the desktop app being in the right state. Local file, browser, and computer access require the app to stay open, so a laptop that sleeps or an app that is closed can pause the work that depends on your machine. And because Cowork needs a paid plan, and on Team or Enterprise possibly an admin to enable it, it is not something every person in every workplace can switch on today.
Finally, Cowork is a poor fit for tasks that need no files or apps. If you want to think through a decision, rehearse a conversation, or ask a general question, a regular chat is simpler. Cowork pays off when the task has inputs on your disk or in your tools and an output that should end up as a file or a message.
Who it is for
Who gets the most from Claude Cowork, and where to start
Cowork is aimed at people whose work is mostly documents, spreadsheets, messages, and tickets: operations managers, analysts, marketers, project managers, consultants, researchers, and anyone who spends hours each week assembling things from other things. If your day includes turning a pile of files into a short summary, or a channel full of messages into an update, this is the shape of work it was made for.
It is not the same thing as Claude Code, which is built around software development in a coding environment. If you write and maintain code, see our guide to Claude Code. If your work is the paperwork around the code, or any other kind of knowledge work, Cowork is the more natural fit, and it connects to GitHub as one of its possible connectors.
The suggested way to begin is to pick a real task you would otherwise do yourself. Anthropic's getting-started guide gives three examples: generate a one-page brief from a folder of PDFs, draft a status update from a Slack channel's recent messages, or clean a messy spreadsheet and chart the results. Authenticate the connectors you use most, such as Slack, Gmail, Google Drive, Asana, or GitHub, and start in Manual mode so you can see how Claude approaches the work.
After a few runs, move trusted routine tasks to Auto, write your global instructions, and set up a project for anything you do repeatedly. The skill that matters throughout is describing work clearly, and that is learnable. Our prompt engineering lessons and AI for beginners path cover the habits that make these requests work.
Learn to hand real work to Claude
Reading about Cowork and being good at Cowork are different things. The free Nightschool AI curriculum is hands-on from the first lesson, so you practice writing the kind of clear briefs that make Claude useful on real tasks.
Frequently Asked Questions
What is Claude Cowork?
Claude Cowork is a mode in the Claude desktop app, on macOS or Windows, that works on tasks using your own files and connected apps. Instead of only answering in a chat, Claude plans and carries out steps, reading and writing files in folders you connect. You choose how often it pauses for approval, and it always asks before permanently deleting files.
Do I need a paid plan to use Claude Cowork?
Yes. Cowork requires the Claude desktop app and a paid plan: Pro, Max, Team, or Enterprise. On Team or Enterprise plans, an admin may also need to enable Claude Cowork for the organization before a member can use it. If the mode is missing for you at work, ask your workspace admin first.
Can Claude Cowork see all the files on my computer?
No. Claude sees no files until you explicitly connect a folder. You are prompted the first time you ask it to work with files, and you can connect more folders later. A good habit is to connect the specific folder a task needs, such as one project folder, rather than a broad location that holds unrelated personal or work files.
What are Manual, Auto, and Skip modes in Cowork?
They are permission modes you change from the mode selector in the chat box. Manual pauses for your approval before each action. Auto keeps working while a safety review checks each action for risks like data exfiltration or prompt injection and blocks unsafe ones. Skip pauses and checks nothing, and Anthropic says to use it only when you completely trust every action, connector, file, and app involved.
Can Claude Cowork delete my files?
Not without your say-so. Claude Cowork always requires explicit permission before permanently deleting any files. You see a prompt and must choose Allow. This applies regardless of permission mode. You can still reduce risk further by working in Manual mode and asking Claude to save changes as new copies rather than overwriting originals.
Does Cowork keep running if I close the app?
Partly. Long Cowork tasks can keep running in the cloud when the desktop app window is closed. However, local file, browser, and computer access require the desktop app to stay open. So a task that depends on your connected folders needs the app running, while work that does not touch your machine can continue without it.
Where does Cowork run code and shell commands?
Shell commands and generated code run in an isolated environment on Anthropic's servers, separate from your own computer and network. Reading and writing files in connected folders is a separate matter: on desktop, Claude reads and writes local files in those folders directly, without manual uploads or downloads.
What connectors can Cowork use?
Cowork works with MCP connectors. Anthropic's getting-started guide names Slack, Gmail, Google Drive, Asana, and GitHub as examples to authenticate if you use them most, and says Claude may suggest a connector when a task calls for one. Each connector's permission can be Always allow, Needs approval, or Blocked, and a blocked tool is denied in every mode.
How do I delete a Cowork task, and what happens to the data?
Deleting a task removes it from your history immediately and from backend storage within 30 days. If you do not want a task to use or build memory, you can turn memory off for that single task before its first message. That setting cannot be changed once the task has started, so decide before you send anything.
What is the difference between Cowork and Claude Code?
Cowork is a mode in the Claude desktop app for tasks that use your files and connected apps, such as briefs, updates, and spreadsheets. Claude Code is a separate tool aimed at software development work. If your job is mostly documents, messages, and data, start with Cowork. If it is mostly writing and maintaining code, start with Claude Code.
What should my first Cowork task be?
Pick a real task you would otherwise do yourself. Anthropic's getting-started guide suggests generating a one-page brief from a folder of PDFs, drafting a status update from a Slack channel's recent messages, or cleaning a messy spreadsheet and charting the results. Start in Manual mode so you can see each action, and ask for new copies of files instead of edits.
Ready to put Claude to work?
The best way to understand a coworker is to give them something real to do.