Getting Started with Claude Cowork

1. Installation and Initial Setup

Getting Claude Cowork running on your machine is a three-step process: download the application, connect it to your account, and grant the system-level permissions it needs to operate on your desktop. Each step is straightforward, but the third — granting permissions — is where most new users spend the most time, and understanding why those permissions are necessary will help you make good decisions about what to allow.

Download the Cowork desktop application from your organization’s software portal or directly from anthropic.com/cowork if you are on a personal enterprise plan. The installer is a standard macOS or Windows application package; there is no command-line setup required. Once installed, Cowork will prompt you to sign in with your Anthropic account or, if your organization has configured SSO, with your company credentials. This sign-in step ties your Cowork session to your organization’s permission policies, which your IT administrator may have already configured before you received the invitation to install.

After sign-in, Cowork will request two categories of system permission. The first is screen access: Cowork needs to be able to see what is on your screen in order to understand the state of the applications you are asking it to work in. The second is file system access: Cowork needs to read and write files to actually perform edits. Both permissions are granted through your operating system’s standard privacy settings panel — on macOS this is System Settings → Privacy & Security, on Windows it is through the Accessibility and File System permission dialogs. You can scope the file system permission to specific folders if you prefer to start with a narrow access footprint, which is often a good idea.

2. Understanding the Permission Model

Cowork is designed around explicit consent, and that design philosophy runs all the way through the interaction model. Before Cowork takes any action that modifies something — editing a file, sending an email, submitting a form — it will ask for your confirmation unless you have explicitly told it that operation is always allowed. This is not an obstacle to productivity; it is the mechanism that keeps you in control while you are building trust with a new tool.

The permission model has two layers. The first is the session-level prompt: Cowork asks “may I do X?” and you can say yes (for this instance), yes always (add it to the trusted operations list), or no. The second layer is the trusted operations list, which you build up over time. If you find yourself repeatedly approving the same type of action — say, editing files in your Projects folder — you can add a rule that says Cowork may always perform that class of action without prompting. This is the “always allow” mechanism, and it is how Cowork shifts from feeling like a supervised assistant to feeling like a genuinely autonomous workflow partner.

Enterprise deployments may have this list partially pre-configured by your administrator. If your organization has decided that Cowork is permitted to edit documents but not send emails without approval, those constraints will be enforced at the policy level and will override any rules you set locally. This is by design: centralized policy is what makes Cowork safe to deploy at scale, where individual users cannot always be expected to configure permissions optimally.

3. How to Describe Tasks Effectively

The most important skill in using Cowork well is learning to describe what you want clearly. The model is highly capable, but it works best when you give it a specific outcome to aim for rather than a vague direction. The difference between “fix my spreadsheet” and “in Q2 Budget.xlsx on the Finance tab, update the formula in column D so it uses the revised tax rate of 28% instead of 24%” is the difference between a result you have to review extensively and one that is immediately usable.

A good Cowork task description has three components: the target (which file or application), the current state or problem (what is wrong or what needs to change), and the desired outcome (what done looks like). You do not need to tell Cowork the steps to take — figuring out the steps is its job. Telling Cowork to “open the file, find column D, select the cells, edit the formula” is not more helpful than describing the outcome; it is often less helpful, because it constrains Cowork to your mental model of how to do the task rather than letting it use the most efficient approach it has available.

Application and file references work best when you are specific. “My document” is ambiguous; “the Q3 report in ~/Documents/Reports/Q3-2026.docx” leaves no room for confusion. If you are working in a browser, name the site and the page. If you are working in a tool like Slack, name the channel or conversation. The more precisely you anchor the task to a specific location, the less time Cowork spends figuring out what you are pointing at and the faster it gets to work.

4. What Cowork Can Work With

Cowork’s computer use capability is broad, but it is worth knowing what falls squarely within its design envelope versus what is experimental or outside it. The core supported use cases at GA are documents (Microsoft Word and macOS Pages files, PDFs opened in a reader, Google Docs accessed through a browser), spreadsheets (Excel and Numbers files, Google Sheets in a browser), email (composing and replying in Outlook desktop, Apple Mail, and Gmail in Chrome), messaging (Slack desktop app, Teams, and browser-based chat tools), and general browser navigation for research and form-filling tasks.

File formats matter: Cowork reads and edits the files as they exist on disk, which means it works best with formats it can parse. Proprietary binary formats with heavy macros may produce unexpected results. Plain text files, markdown, CSV, and standard Office XML formats (the default save format for modern Word and Excel) work reliably. PDFs are read-only — Cowork can extract content and summarize, but it cannot write back to a PDF without converting it first.

What Cowork cannot do at GA is equally important to understand. It cannot run arbitrary code or scripts on your behalf. It cannot access files on network drives that are not mounted locally. It cannot operate in virtual machines or remote desktops. And it cannot act inside applications that use non-standard rendering pipelines (some Electron apps with custom accessibility implementations may have limited support). If you try a task and Cowork reports it cannot see the relevant interface element, this is usually why.

5. Your First Tasks and Best Practices

The best first tasks for Cowork are ones where you already know exactly what the answer should look like. This lets you verify the output quickly and build calibration for what Cowork does well before you assign it to tasks where you are less certain of the expected result. Good starting tasks include: summarize a specific document in three bullet points, update a specific cell formula in a spreadsheet, draft a reply to a specific email in your inbox, and pull a table from a webpage into a new spreadsheet file.

When Cowork is about to edit a file, it will show you a diff view — a before-and-after comparison of the changes it plans to make — before writing anything to disk. This is the moment to catch mistakes. Read the diff carefully for your first few tasks, even if the task was simple. Understanding what Cowork changed and how it interpreted your instructions will help you write better prompts for subsequent tasks. Once you have seen enough diffs to trust Cowork on a certain type of task, you can move faster by approving quickly. Until then, the diff is your safety net.

Keep Cowork focused on one task at a time. Unlike a human assistant who might be able to juggle multiple simultaneous requests through mental context switching, Cowork performs best when it completes one task fully before starting the next. Giving it a list of five things to do at once can cause it to lose track of context, apply an edit from task three to the file it was working on for task one, or ask clarifying questions mid-run that interrupt your other work. Sequence your tasks and let each one complete.

🧠 Check your understanding

5 questions · your answers are saved in this browser only

1. 1. What system permissions does Cowork require to function, and why?

   Microphone access only, to listen for voice commands Screen access (to see application state) and file system access (to read and write files) Administrator privileges to install browser extensions automatically Network access to proxy all your traffic through Anthropic servers

   Cowork needs screen access to understand what is displayed in the applications you ask it to work in, and file system access to actually read and edit your files.

2. 2. What is the 'always allow' feature in Cowork's permission model?

   A setting that lets Cowork act without ever asking for confirmation on any task A rule you add to the trusted operations list so Cowork can perform a specific class of action without prompting each time A mode where Cowork only reads files and never writes An enterprise policy that disables all permission prompts organization-wide

   Always allow lets you add specific operation types to a trusted list, so Cowork can perform those actions automatically while still prompting for anything outside that list.

3. 3. Which of these task descriptions would give Cowork the clearest direction?

   'Fix my spreadsheet' 'Update the document about the project' 'In Q2 Budget.xlsx on the Finance tab, change the tax rate in column D formulas from 24% to 28%' 'Make the email sound better'

   The most specific description names the file, the location within it, and the exact change to make. This leaves Cowork no ambiguity about what to do or where to do it.

4. 4. What does Cowork show you before writing any changes to a file?

   A natural-language summary of what it plans to do, with no code or file content A diff view showing the before-and-after comparison of the planned changes A progress bar while it writes the file in the background A notification after the save is already complete

   Cowork shows a diff — a before-and-after comparison — before writing to disk. This is your opportunity to catch mistakes before they are committed to the file.

5. 5. Why does Cowork perform better when given one task at a time?

   It is technically unable to process more than one file simultaneously Running multiple tasks simultaneously drains the battery faster Juggling multiple tasks can cause context confusion, leading to edits being applied to the wrong file or mid-task clarifying questions Enterprise licenses only permit one active task per session

   Multiple simultaneous tasks can cause Cowork to lose track of context — applying an edit meant for one task to another file, or interrupting your workflow with clarifying questions mid-run.

Reset quiz

🛠️ Build it yourself

Follow these exact steps to reproduce it yourself

• Day 1: Complete installation, grant screen and file permissions scoped to Documents and Downloads, and sign in with your organizational account
• Day 1: Run your first task — ask Cowork to summarize a document you already know well, then compare its summary to your own understanding to calibrate accuracy
• Day 2: Try a spreadsheet task — give it a specific cell or column to update and review the diff carefully before approving
• Day 3: Draft an email reply using Cowork — give it the thread context and the key points you want to make, then edit its draft before sending
• Day 4: Identify one recurring workflow (a weekly report update, a routine data pull, a templated document) and use Cowork to do it end to end
• Day 5: Review your permission settings — add trusted operation rules for the actions you approved repeatedly this week, and tighten any permissions that felt too broad
• Week 2: Begin chaining tasks: after Cowork finishes one task, immediately assign the next in the same session to build familiarity with its session memory