How to Use Claude Cowork (September 2026)

Here is the short answer to how to use Claude Cowork. You open the message box in Claude, select "Cowork" in the bottom left corner, describe the outcome you want instead of the next step you want, choose how much approval to give, and come back later to finished files you can preview and download. Claude runs the task in the cloud, breaks it into subtasks, runs code and shell commands in an isolated environment, and delivers the output back into the session, as set out in Anthropic's Cowork guide.

This guide takes you from a paid Claude account with nothing set up, through one complete real task with numbered steps, to that same task running on a weekly schedule without you touching it. It also covers what Cowork can reach on your computer, what it cannot do yet, and what Claude Code users are actually getting that is different.

What Claude Cowork is

Claude Cowork uses the same agentic architecture that powers Claude Code, with no terminal required. In a normal conversation, Claude responds to one prompt at a time and waits for you. In Cowork, Claude takes on a complex multi-step task and executes it. You describe an outcome, step away, and come back to finished work.

When you start a task, the sequence is consistent: Claude analyzes the request and creates a plan, breaks complex work into subtasks, runs code and shell commands in an isolated environment, coordinates parallel workstreams where appropriate, then delivers finished outputs to the session where you preview and download them.

Sessions run in the cloud on Anthropic's servers, and both the sessions and the files are saved to your Claude account, across every surface. That is why closing your laptop does not kill the run. When a task needs something that lives on your machine, such as a local file or your browser, Claude reaches it through the Claude Desktop app on that computer.

Who can use Claude Cowork right now

Cowork is on paid plans only. Free accounts do not have it.

If you are on Team or Enterprise and something described here is missing from your account, an admin setting is the usual reason. Admins on those plans also control whether the automatic approval mode is available at all, and an organization can require per-task approval for connector tools that can write or delete.

What you need before you start

Three things, and only the first is non-negotiable for every task. A paid subscription. For local file access, browser use and computer use, the Claude Desktop app for macOS or Windows, open and connected on the computer holding those files. An active internet connection for the whole session.

The second one causes most of the confusion people run into later. A cloud session can read and write files in connected folders only while the desktop app is open on that computer. If the app is closed, the session keeps running, but it cannot reach anything local.

How to open Claude Cowork on each surface

On desktop, web and mobile, chat and Cowork share one home. Find the message box, select "Cowork" in the bottom left corner, and describe the task. To go back to a regular conversation, select "Chat" in the same place.

The Chrome side panel works differently. Opening the panel starts a Cowork session directly, with no selector to switch.

Sessions are portable. You can start a task on one surface, open the same session from another to check progress, answer a question Claude has asked, or redirect the work, then pick up the finished output wherever you are.

How to set global instructions before your first task

This step takes two minutes and improves everything that comes after it, because it stops you re-explaining who you are in every prompt.

1. Open Settings, then Cowork.
2. Click Edit next to Global instructions.
3. Type the context Claude should carry into every session. Preferred tone, the output format you want by default, and background on your role are what this field is for.
4. Save.

Global instructions apply to every Cowork session. There is a second layer for context that applies to one body of work: folder instructions, which you get when you select a local folder on desktop. Those hold project-specific context, and Claude can update them itself during a session as it learns how you want things done.

One current limitation is worth knowing before you write these. What Claude remembers about you in chat does not carry into Cowork yet, and inside Cowork memory is supported in projects only. Global instructions are doing the work that chat memory does elsewhere.

What the three permission modes do

You pick a mode from the mode selector in the chat box, and you can change it at any time.

Manually approve, previously called "Ask before acting." Claude pauses and asks for approval before acting, and you choose Allow or Deny. Use this for your first few runs, because watching what Claude asks permission for teaches you what it is actually doing.

Automatically approve. Claude keeps working without asking at every step, while reviewing each action for safety concerns such as data exfiltration and prompt injection, and blocking anything it determines to be unsafe. When something is blocked, Claude looks for a safer route or asks you directly, and if it keeps hitting blocks it goes back to asking at each step. Two practical notes: this mode consumes more of your usage limit than the others, and Anthropic advises staying close for work involving money, messages sent as you, or important files. No mode replaces your own judgment about what you handed over.

Skip all approvals, previously called "Act without asking." Claude does not pause, and nothing checks the actions automatically. This is only for tasks where you trust every element involved.

Connector permissions sit on top of these. A connector tool set to "Always allow" is approved in Manual and Skip modes, and in Auto its read-only tools are approved while Claude decides on write and delete actions. A tool set to "Needs approval" asks you in Manual, is decided by Claude in Auto, and is approved in Skip. A tool set to "Blocked" is denied in all three modes.

One thing is constant across every mode: Claude requires explicit permission before permanently deleting any files, and you have to select "Allow" on that prompt.

How to run your first Claude Cowork task, step by step

The task below is a folder of raw exports on your machine turned into a cleaned, formatted deliverable. Receipts into an expense report is the version written out here. The same shape covers a folder of CSV exports into a summary spreadsheet, a set of transcripts into a themes and action items document, or rough notes into a slide deck.

1. Put the raw files in one folder on your computer. It does not need to be tidy. Mixed file types are fine, and messy filenames are fine, because renaming can be part of what you ask for.
2. Open the Claude Desktop app on that computer. On Windows, update it first if you have not recently, since Cowork requires the latest version. Leave the app open for the whole task. If your computer sleeps or the app closes during a task that needs local files, that is the most common reason a task stops mid-run.
3. Select "Cowork" in the bottom left corner of the message box.
4. Set the mode selector to Manually approve for this first run.
5. Select the local folder holding your files, so the session can reach it.
6. Describe the outcome in one message, not a sequence of small instructions. Something in this shape:

Two things make that prompt work. It names the finished artifact, and it says what to do with the cases Claude cannot resolve. Ambiguous items become a visible list you check in a minute rather than silent guesses buried in a total.

7. Watch the first minute. Progress indicators show what Claude is doing at each step, and Claude surfaces its reasoning so you can follow along. If the plan is reading the folder in a way you did not intend, jump in to course-correct while it runs. You do not have to stop and restart.
8. Answer the approval prompts. In Manual mode you get Allow and Deny choices. Reading them is the fastest way to understand what the task actually touches.
9. Let it run. If Claude is coordinating multiple sub-agents in parallel, the progress view shows more than one thing happening. You can close the laptop lid at this point only if the task no longer needs the local folder. While local files are involved, keep the desktop app open.
10. Preview the output in the session, then download it. Claude names the output location, so if the file is not where you expected, that is the first place to look.
11. Ask for changes in the same session. Wrong category mapping, a tab you want split, formatting you want changed. For a Markdown draft Claude produced, you can highlight the text you want changed, click "Edit with Claude," and type your request, so the edit lands exactly where you marked it. Spreadsheets and slides can be taken further with Claude for Excel and Claude for PowerPoint.
12. Store what you learned. Ask Claude to write the category rules, the naming convention and the tab structure into the folder instructions. Next month's run starts from your rules instead of from scratch.

Expect the first run to take longer than you think and to land close but not finished. The second run, with folder instructions in place, is usually where this starts saving real time.

How Claude Cowork scheduled tasks turn that into a weekly job

Once one run works, you turn it into a recurring one. There are two ways in: type /schedule inside any Cowork task, or click "Scheduled" in the left sidebar, where you can view, create and manage scheduled tasks.

Scheduled tasks run in the cloud, so they do not need your computer awake or the desktop app open. That is the whole point of them, and it is also the constraint you have to design around.

Here is the practical consequence for the expense report task. A scheduled cloud run cannot open a folder that sits on a closed laptop. You have two clean options.

Option one: keep the inputs local and accept the dependency. The desktop app has to be open on that computer when the schedule fires. Workable if your machine is on during the working day and the run is scheduled for a weekday afternoon.

Option two: move the inputs somewhere the cloud can always reach. If the receipts land in a Google Drive folder or arrive by email, a connector gets Claude to them without your machine being involved at all. This is the version that keeps working while you are on holiday.

When you set up the recurring version, rewrite the prompt so it is self-contained. A scheduled task has no one sitting there to answer a question. State the date range in relative terms, such as the previous seven days rather than June and July. State where the output goes. State what to do with items it cannot resolve, since nobody will be there to click Allow on a judgment call.

Mode choice matters more here too. A scheduled task in Manual mode will sit and wait for approvals you are not there to give. Auto mode is the usual fit for a schedule you trust, keeping in mind that it uses more of your usage allocation, and that on Team and Enterprise your admin decides whether it is available.

Where Claude Cowork can and cannot reach your files

Local connectors and plugins that contain local MCP servers are desktop only. Local file access, browser use and computer use from web and mobile work only while the Claude Desktop app is open on your computer. Close it and the session continues, but the local folder goes out of reach.

How Claude Cowork chooses between connectors, your browser and your screen

For tasks that touch other applications, Claude uses the most precise tool first. Connectors come first, such as Gmail, Google Drive and Slack, because that is the fastest and most reliable path. When there is no connector for what you need, Claude navigates the Chrome browser through Claude in Chrome. Screen interaction comes last, where Claude clicks, types and navigates your desktop apps directly, because it takes much longer and is more error-prone.

If a task keeps failing or crawling, a missing connector is often why, and adding one converts screen-clicking into a direct call.

Computer use has no sandbox between Claude and your applications, so it carries specific safeguards. Permissions are per application, Claude asks before accessing each one, some sensitive apps including investment and trading platforms and cryptocurrency are blocked by default, and you can add your own apps to a blocklist. Your computer has to be awake with the desktop app open for any of it.

How to publish your own expertise as a connector Cowork can call

Everything above treats connectors as things you install. They are also things you can publish, and that is the part most people miss when they first understand how Cowork works.

When Cowork hits a decision it cannot make well on its own, whether that is which pricing structure to recommend, whether a contract clause is standard, or how to program around an injury, it answers from general training. A connector lets it consult a specific person instead. If you are the person people already ask, that consultation is something you can charge for.

That is what Callstand does. You publish an expert connector and get a handle at callstand.com/p/yourhandle. It exposes three verbs: a plan before work starts, a finished deliverable written in your style, and a review of work before it ships. You set your own price for each verb, and buyers hold prepaid credits with per-call spending caps, so there is no invoicing and no scheduling on either side.

Because it runs as an MCP server, Cowork and the rest of Claude can call it directly. Someone running a scheduled task in Cowork can have your judgment checked into the output before anything is delivered, at 2am, without you being awake or involved.

What building one actually involves

No prompts to write, no model to train, no code, and no infrastructure. A guided setup captures how you think about the work by having you react to examples rather than write rules, which is the part most experts find they can do easily and quickly. Most people finish in an afternoon. You then choose who can call your connector, approve buyers individually if you want to, and your material stays private, since callers receive answers rather than access to what shaped them.

There is no follower count or audience size requirement. The bar is having judgment other people would rather borrow than reconstruct.

How to keep recurring work organized with projects and plugins

Projects group related tasks into separate workspaces with their own files, context, instructions and memory. They are available on every surface, and from inside a project you can start either a chat or a Cowork session, with Claude using the project's knowledge as context. Since memory inside Cowork currently exists in projects only, a project is where ongoing work belongs rather than a string of one-off sessions.

Two details to plan around. Projects tied to a local folder support Cowork sessions on desktop only, and Cowork will not change a project's contents, so anything you want kept has to be added to the project yourself.

Plugins bundle skills, connectors and sub-agents into a single package that can be customized for a role, a team or a company. If several people do a version of the same monthly task, that packaging is what stops each of them rebuilding it.

How much of your usage limit Cowork uses, and what it cannot do yet

More than chatting, and noticeably so, because multi-step tasks are compute-intensive and require more tokens to execute. Automatic approval mode uses more than the other modes.

Anthropic's own guidance is to batch related work into single sessions, use standard chat for simpler tasks that do not require file access or extended execution, and monitor your usage in Settings then Usage. Hitting limits faster than you expect is normal behavior rather than a fault.

On the gaps: what Claude remembers about you in chat does not carry into Cowork sessions yet, and within Cowork memory is supported in projects only. Sessions cannot be shared with others, though Team and Enterprise users can share live artifacts within their organization. Live artifacts and plugins that include local MCP servers work through the desktop app only.

How to fix the most common Claude Cowork problems

"Setting up Claude's workspace" appears at the start of a session. This is expected and means Cowork is updating to the most recent version to apply fixes and improvements. Wait it out.

Claude stopped partway through a task. For local sessions, check the desktop app stayed open the entire time, since a closed app or a sleeping computer can end the session. Cloud sessions keep running in the background and can be reopened from any surface.

You hit your usage limit quickly. Expected with multi-step tasks. Batch work into fewer sessions and keep simple requests in chat.

The output file is not where you expected. Check that you granted the appropriate file access permissions, then check the output location Claude specified.

You want a task gone. Use the "⋮" next to the task, or select tasks from your Tasks list and click the trash icon. It leaves your task history immediately and backend storage within 30 days.

For Team and Enterprise administrators: Cowork via mobile and web is captured in the Compliance API, activity can be monitored with OpenTelemetry, and network egress permissions are respected, although they do not apply to web fetch, web search, or MCPs including Claude in Chrome. Web fetch runs server-side and is limited to search results and URLs you have shared. Owners can turn off web search for Cowork and Chat under Organization settings, then Capabilities.

Claude Cowork vs Claude Code

Same agentic architecture underneath. Cowork removes the terminal.

If you already live in Claude Code, what Cowork adds is the surface area around the agent rather than a different agent. Sessions you can start on desktop and check from your phone. Scheduled runs in the cloud that do not depend on your machine being awake. Connectors to Gmail, Drive and Slack as first-class tools. Outputs that arrive as Excel files with working formulas and PowerPoint decks. Permission modes exposed as a selector rather than flags.

If you are not a developer, the honest framing is that Cowork gives you the same execution engine engineers have been running, in a place where you never see a command line. The skill it asks for is describing outcomes precisely and deciding how much approval to hand over, and both are learnable in an afternoon.

Where Claude Code still fits better is anything anchored in a repository and a terminal workflow. The two are not competing for the same hour of your day.

Frequently asked questions

Changelog: published 25 August 2026; reviewed 1 September 2026. Availability, surface support and permission mode behavior verified against Anthropic's help center on the later date.