Claude Cowork is an agent in Claude that carries out tasks on its own. Cowork started in the Desktop app and has also worked on the web and mobile since July 2026. Since 6 October 2026, new tasks on Pro, Max and Ultra have run in the cloud by default.
It does this not by chatting about the work, but by reading, editing and creating files, or by running a task in the cloud that does not depend on your local files. You describe what needs to be done, and Cowork works through it while you get on with something else.
The step from chat to agent is bigger than the name “Cowork” suggests. A chat answers your question; an agent carries out a task.
That difference decides when Cowork adds value and when an ordinary Claude chat is enough.
How does Claude Cowork work in practice?
Claude Cowork works by getting access to files, folders and connectors, and then running a task from start to finish in a virtual environment. Cowork is the name Anthropic uses for Claude’s agent mode.
Moving to an agent only pays off once you have the basics in place in the chat.
Cowork does not run locally on your computer by default, even if it feels that way. Tasks involving local files, browser use or computer use need the Desktop app on macOS, Windows or Linux (beta), but the processing itself always happens in Anthropic’s cloud. For tasks that do not depend on your local drive, a session can also run entirely on the web or mobile.
Instead of question and answer, Cowork works towards an end goal: you give it a task, it draws up a plan, carries out subtasks and delivers the result.
Reading and writing local files, building documents automatically, visiting websites and collecting data through connectors all happen within one session, without you having to step in at each stage.
Cowork is built on the same agentic architecture as Claude Code, the terminal tool for developers, but without the terminal. Anthropic has made that way of working accessible to knowledge workers without a technical background.
The aim is not to be a better chatbot, but to get repetitive work finished while you do something else.
How is Cowork different from an ordinary Claude chat?
What sets Cowork apart from an ordinary Claude chat comes down to three things: the material Cowork works on, the length of a session, and the kind of output it produces.
Cowork works on your real files, it carries out several steps in a row without you confirming each one, and it delivers not text but files, folder structures and completed actions.
In an ordinary Claude chat you upload a file and get an answer in the chat. What you do with that answer next (formatting a document, filling in a spreadsheet, putting together a report) is up to you.
In Cowork you describe the end goal and Claude handles the whole process. It opens the file, reads the content, writes a new document to your hard drive, saves it in the right folder and, if needed, repeats that for ten or a hundred other files.
The second difference is the length of a session. A chat is by definition a conversation: you ask something, Claude answers, done. Cowork supports long-running tasks without the session timing out.
Since July 2026 Cowork has also run on the web and mobile, with a session running entirely on Anthropic’s servers: you start a task on your laptop, close the lid, and Claude carries on without any of your devices needing to stay on.
Since February 2026 there have also been scheduled tasks, where Cowork carries out work periodically without you having to start it again.
The third difference is the output itself. A chat gives you text; Cowork gives you tangible results: a spreadsheet you open in Excel, a Word document sitting in your folder, an organised folder in which all your loose files have been sorted into the right subfolders.
If you lose a lot of time on the step “now put this text in the right format and save it in the right place”, that is a very different proposition from a chatbot.
The difference between Cowork and an ordinary Claude chat has been shifting since 16 September 2026. Anthropic is rolling out a new Claude experience in which everything Cowork does is available from any conversation, in phases for Pro and Max on the web, desktop and mobile. You create designs, slides and docs in any conversation, and Artifacts have since been available on all plans, including the free plan. The distinction then mainly lies in what you ask for: an answer, or a task carried out on files.
Where Cowork edits files and carries out tasks, the design tool Claude Design focuses on creating visual work in its own canvas.
For a single user that distinction is quick to grasp, but as soon as a whole team starts using it, it is something you need to explain once and write down, rather than letting everyone work it out for themselves. That handover is the core of AI implementation within a team.
What do you use Claude Cowork for?
Cowork is good at tasks that consist of many small steps, that work with local files, and that end in a tangible deliverable. Four categories keep coming back in the use cases Anthropic and early users have shown.
The first category is organising files at scale. Cowork can sort a messy Downloads folder with hundreds of files by type, date or content.
It recognises that a PDF is an invoice and puts it in an Admin folder; another PDF that turns out to be a manual ends up in Documentation.
Doing that by hand takes an hour; Cowork does it in minutes and produces a log of what was moved.
The second category is reporting from scattered sources. Combining meeting notes, client briefings, loose memos and source files into one structured document is exactly the type of work Cowork is good at.
The added value is not in the quality of the writing (an ordinary Claude chat does that just as well) but in the assembly: finding the right material, putting it in the right order, formatting it and saving it in the right place.
The third category is data analysis on local spreadsheets and datasets. Cowork can open an Excel file, create pivot tables, add new sheets with the results of the analysis, and generate charts.
The output is a completed file you open straight away in Excel, not a written description of what it ought to contain.
The fourth category is repeatable work through Cowork projects. Since early 2026 Cowork has supported its own projects with persistent instructions, a dedicated folder and scheduled tasks.
That lets you set up a weekly report, a recurring email check or a periodic data update as a structured workflow, so the same task starts automatically at a fixed time.
How do you get started with Claude Desktop?
The Claude Desktop app is available on all plans, including the free plan, for macOS and Windows (including arm64) and in beta for Linux (Ubuntu 22.04 or later, Debian 12 or later). Cowork itself requires a paid plan: Pro, Max, Ultra, Team or Enterprise.
Cowork originally ran only through the Desktop app, but since July 2026 it has also worked on the web and mobile: on Pro, Max and Ultra, on Team in beta, and on Enterprise once an owner switches it on. For tasks involving local files, browser use or computer use, the Desktop app remains the full version.
You download the Claude Desktop app from claude.com/download; the guide to the Claude desktop app covers installation for each operating system. Cowork works on macOS, Windows and Linux (beta), provided you use a recent version of the app.
After installing, you sign in with your Claude account and select Cowork in the message box, provided you are on a paid plan. If your message box shows no choice between Chat and Cowork, you have the new Claude experience that has been rolling out to Pro and Max in phases since 16 September 2026, and any conversation can pick up a Cowork task.
The next step is choosing a working folder. Cowork only has access to folders you explicitly share. That is a deliberate design choice: the system should not have free access to your entire hard drive.
For a first test, it makes sense to create a separate folder (for example “Cowork-test”) and put a few sample files in it. Start small and only widen the scope once you know what the agent does in your day-to-day work.
The third step is the first task. Describe the end goal in plain language, not the steps Claude has to take.
“Read all the PDFs in this folder, identify which ones are invoices, and put them in a subfolder called Invoices with the supplier’s name in the file name” works better than a list of micro-instructions. Cowork plans the steps itself; your job is to state the end goal as clearly as possible.
How do you give Cowork good instructions?
Good instructions for Cowork differ from good prompts in an ordinary chat. A prompt asks for a written answer; a Cowork instruction describes an end result with enough context for Cowork to build it on its own. Three principles keep coming back.
The first principle is outcome over execution. Describe the result you want to see, not the exact steps Claude should take.
“Create a spreadsheet of all the expenses on these receipts, grouped by month and by category, with a total column” leaves room for an efficient approach. A rigid step-by-step instruction (“open each file, read the amount, copy it to row X”) forces a way of working that may not be the best one.
The second principle is setting out context in advance through instructions. In Cowork’s settings you can set global instructions that apply to every session, and folder-specific instructions that are only active within a particular folder.
For recurring work, that pays off. A “Client reports” folder with instructions such as “always put an executive summary of five bullets at the top, use our house-style headings” gives consistent output without you having to write it out again for every task.
Stating a role in those instructions helps as well.
The third principle is building in checkpoints for high-risk tasks. Cowork can be set to ask for confirmation before certain actions (for example deleting a file, sending an email or carrying out a transaction).
For irreversible actions on real data, that is not a needless step. It costs you an extra click per action and gives you the chance to correct Claude before a wrong assumption causes damage.
What are the limitations of the desktop version of Claude?
Since October 2026 the limitations of the desktop version of Claude mainly concern work that needs your own computer. Cowork itself is less tied to the desktop than it was at launch. What you still need to watch is which work is tied to your computer and whether Cowork’s logging meets your organisation’s requirements.
Running Claude locally is only partly possible: access to your files and browser control happen on your own computer, but the processing always runs through Anthropic’s servers. There is no offline mode, and Cowork always needs an active internet connection.
Since 6 October 2026 new Cowork tasks on Pro, Max and Ultra run on Anthropic’s servers by default, and the option to run a task only on your own computer has gone. The same sessions and files are available on desktop, web and mobile: start on your work laptop, check progress from your phone, pick up the result on your home computer. On Enterprise, an owner first has to switch on Cowork in the cloud.
Tasks that need local files, local connectors, browser control or computer use do remain tied to your computer. The desktop app then has to stay open on the machine where those files and apps live. Tasks you had already started on your computer are finished there, and scheduled tasks that were already running locally keep running there. If you lose your connection during one of those long local tasks, the session can end up in an inconsistent state.
Governance has been extended since launch, but you still need to assess it yourself. According to Anthropic, Cowork activity on the web and mobile is recorded in the Compliance API, with each session traceable to a user. Anthropic does not say whether local desktop sessions are included there too. Team and Enterprise owners can also send Cowork events to their own SIEM or monitoring tools through OpenTelemetry. That event stream is intended for monitoring and runs alongside the Compliance API; it does not replace that record.
Organisations in financial services, healthcare or government, where every action has to be traceable, should check in advance whether that logging is sufficient.
From my years in the banking sector, I see this point clearly: a tool that does not produce an audit trail is not usable for processes that fall under supervision, however much value it adds otherwise.
So the question is no longer whether Cowork records anything, but whether that record covers what your regulator or internal audit asks for, including for tasks that run locally on a computer.
Cowork’s results transfer easily; the way of working does not. The files Cowork produces are ordinary Word, Excel and PDF files you can open anywhere.
New scheduled tasks run in the cloud and global instructions apply to every Cowork session in your account; only folder instructions belong to a locally connected folder. If you do not share those instructions with your team in a structured way, the way of working stays with you.
What is the difference between Cowork, Projects, Skills and Artifacts?
Cowork is often confused with three other Claude features: Projects, Skills and Artifacts. They are related, but they solve different problems.
A Claude project in claude.ai is a workspace for chats about one subject, with shared instructions and a knowledge base of uploaded documents. It runs in the browser.
A Skill is a reusable, structured set of instructions (in a SKILL.md format) that Claude calls up automatically when a particular task comes along.
An artifact is not a workspace or a set of instructions but an output format: a block of working output that appears in its own window in claude.ai.
Cowork is an agent that carries out work for you, with its own projects, its own instructions and its own memory structure, on your computer or through a cloud session.
The practical dividing line is what you want to end up with. If you want a series of chats on the same subject with shared context, a claude.ai project is the right place.
If you want to capture a way of working that applies across several subjects, you choose a Skill. If you want a working result in its own window that you can share or remix, you end up with Artifacts.
If you want the work actually carried out, with files as output, Cowork is the feature that fits. The four features do not exclude each other and work well side by side.
Where Cowork works on your files and tasks in the desktop app and in the cloud, Claude Tag brings the same assistant to Slack, so colleagues can call on it in a channel.
Who is Claude Cowork suitable for?
Claude Cowork suits knowledge workers who regularly do structured work with local files and produce a tangible deliverable from it. Researchers, analysts, operations teams, lawyers, finance professionals and marketers are among the groups where Cowork saves the most time.
If you never work with individual files, the added value is limited. Someone who spends the whole day in a browser application or a SaaS tool, and whose output consists of actions inside those tools, gets little back from Cowork.
The agent can reach external systems through connectors, but the core of the product is editing files, on your own computer or in the cloud.
For organisations that want to work with Claude as a team, Cowork is a step that follows setting up projects and agreeing a shared way of working, not a replacement for them.
For SMEs that want to use agent features seriously, Cowork should be covered by fixed team agreements, because access to files calls for clear arrangements about what the agent may and may not touch.
A logical order: first set up a claude.ai project with instructions and knowledge that describe the way of working, then let individual team members use Cowork for the execution, and feed the lessons learned back into the project instructions.
If you are wondering which combination of Claude features fits your work, the quickest way to find out is to decide for each task what you need: an answer, a workspace or a completed assignment.
The Claude Course covers the choice between chat, projects, skills and Cowork in the context of real work scenarios, so you decide which feature to use for what based on the work itself rather than the hype.
The Claude e-learning then covers how to set up Cowork for recurring work and how to stop the feature becoming a toy instead of a productivity tool.
Frequently asked questions about Claude Cowork
Before you let Claude Cowork loose on your own files, you will probably have questions about usage and safety.
Does Claude Cowork use more of your usage limit than a chat?
Yes, Claude Cowork uses more of your usage limit than a short question in the chat, because a task consists of several steps and each step uses tokens. Anthropic does not give a fixed usage figure per task. A long task on a folder with many files therefore weighs more heavily than a single question about one document.
Can Claude Cowork delete files?
Claude Cowork can delete files, but it always asks for your explicit permission before it permanently deletes a file. A prompt appears and the deletion only goes ahead after you choose Allow. According to Anthropic, that confirmation applies in every mode, even if you otherwise let Cowork work on its own. You should also keep a backup of the folder Cowork works in.
Is Claude Cowork safe to use?
Claude Cowork is safest to use with limited access, because according to Anthropic the risk of prompt injection is never entirely zero. In prompt injection, malicious instructions are hidden in a file or website that Claude reads during a task. Anthropic recommends a separate working folder without sensitive files such as financial documents, and the Manually approve mode for actions that are hard to undo.
Does Claude Cowork remember what you discussed in chats?
Yes, in a cloud session Claude Cowork starts with what Claude already remembers from your chats, and information from Cowork tasks comes back into the chat. If you do not want that for a particular task, turn Memory off in the menu behind the + button. Choose this setting before your first message, because it cannot be changed during the task.
