Configure Claude Cowork The complete guide, updated in September 2026

Three months ago, Claude was configured Cowork It used to mean installing a desktop application, opening a folder for it, and accepting that everything would shut down as soon as the computer went to sleep. That's no longer the case. Since the summer of 2026, the sessions Cowork run in Anthropic's cloud, launch from the web, desktop app, phone, or Chrome sidebar, continue running when you close your phone, and remember you from one conversation to the next. Since September 16, 2026, the separation between "chat" and Cowork It even disappears gradually: a single interface, in which Claude decides whether to reply in three lines or to go and work for half an hour on your files.

The principle itself hasn't changed. Rather than responding to one message at a time, Claude takes on entire tasks in several stages: he reads your files, uses your connected tools, and delivers a finished product—a formatted document, a spreadsheet with its formulas, a presentation, a well-organized folder. You describe the expected result, step away, and return to the completed work. If you're looking for the big picture first, our Cowork compare market environments before entering the configuration.

And the observation made by most new users has not changed either: Cowork Out of the box, it's disappointing. Without configuration, the tool doesn't know who you are, how you work, or where to submit your work, and it produces generic results. Properly configured, it's a radically different tool. The difference between the two still comes down to about thirty minutes of setup and a few principles that this article outlines, in its September 2026 version. Once the tool is configured, a well-defined task is the best first attempt, for example... transforming a complex case into an executive presentation or prepare the weekly review of a team.

Article published in June 2026 and completely rewritten in September 2026. The first version described a Cowork executed locally, reserved for the desktop application, with no memory outside of Projects and whose scheduled tasks stopped with the computer; this version integrates the move to the cloud, the arrival on the web and mobile, the new memory, the integrated browser, the three permission modes and the merger with the chat announced on September 16, 2026.

What has changed since June 2026, in one minute

For readers who followed the first version of this guide, here's the gist. The sessions Cowork now run in an isolated and temporary cloud environment, specific to each session, and your tasks and files are saved in your Claude account: a task launched on the office computer can be tracked from the phone and continues to run when the application is closed. Cowork is accessible at claude.ai, in the macOS and Windows desktop application, in the iOS and Android apps (in beta), and in the Chrome sidebar. Memory is no longer limited to Projects: it's account-wide for cloud sessions, with separate memory for each Project. Scheduled tasks run remotely, even when the computer is off. The desktop application now has an integrated browser and direct control of your applications. Permission modes have been increased from two to three. And the setup wizard is now called /setup-claudeThe rest of the article details each of these points.

Cowork or cat: a border that disappears, a rule that remains

Until mid-September 2026, Claude offered two distinct modes. The chat It was a conversation: you wrote, Claude replied in the window, and it was up to you to retrieve the result. Cowork It was a work session: Claude took up the agentic architecture of Claude Code, Anthropic's development tool, without terminal or command line, planned a series of actions, executed them, coordinated several sub-tasks in parallel as needed and dropped real files, to be opened, shared or continued to be edited.

Anthropic observed that its customers "often struggled to choose the right tab for the right task." Therefore, starting September 16, 2026, the two modes have merged into a single Claude experience, rolled out in waves starting with the Pro and Max plans on web, desktop, and mobile; Team and Enterprise organizations will follow, with at least thirty days' notice for the latter. In this new interface, the Chat selector / Cowork no longer exists: you describe the expected result, including its final format, and Claude chooses whether to respond directly or open a work session. Your old tasks CoworkYour projects, connectors, skills, and files are preserved. If your account hasn't yet switched over, you continue to select " Cowork » in the message area.

The distinction remains useful in your mind, however, because it influences the configuration and consumption. Where a chat response tells you what to doa working session the fact Ask someone to organize a photo folder and you get, in one case, a naming convention, in the other, files that are actually organized. The orientation rule is simple: entrust to Cowork Multi-step tasks that involve several files or sources benefit from being executed at length; keep the one-off question and quick reformulation for a direct exchange, which is much less resource-intensive.

The prerequisites: a paid subscription, and the desktop application for everything related to your computer.

Three conditions. A paid subscription, First of all : Cowork This feature is reserved for the Pro, Max, Team, and Enterprise plans and is not included in the free plan. On Team, cloud sessions are enabled by default and an owner can disable them; on Enterprise, it's the opposite: the owner must enable them and then grant access through custom roles.

An access surfaceNext, and now you have a choice: the web (claude.ai), the desktop application for macOS and Windows, the iOS and Android mobile apps, and the Chrome sidebar, which opens directly to a session. Cowork For Max and Team subscribers, and later for Pro subscribers as the rollout progresses, task management, connectors, and scheduled tasks work everywhere. However, anything physically connected to your computer still relies on the desktop application, which must remain open and online: access to local folders, the built-in browser or Chrome extension, control of your applications, and the MCP servers installed on your machine. From a phone or browser, these functions remain available as long as a connected computer with the application open is responding on the other end.

An Internet connectionFinally. It is essential to start and follow the session; but since the session itself runs in the cloud, closing your computer no longer interrupts the work in progress, unless it involves local files.

Where it is executed Cowork : in the cloud, with a bridge to your computer

This is the mental model to acquire before any adjustments, because it explains all the access rules that follow. Each session Cowork is assigned a isolated and temporary cloud environment on Anthropic's servers, destroyed at the end of the session. This is where Claude reads, calculates, writes, and produces his deliverables; this is where your session files, linked to your account, are saved. This environment cannot access your home or company network.

To access your computer, the session needs a bridge: the desktop application. When you link a computer to a session ("Link to this computer"), Claude can read and write to the folders you have explicitly connected, control the browser, and, if you have allowed it, your applications. As soon as the application closes or the computer goes to sleep, the bridge is broken: the session continues in the cloud, but without access to your files, until the link is re-established. The desktop application can also run entirely local sessions, using your machine's resources; note that these local sessions do not have access to your account's memory, which is reserved for cloud sessions.

The practical consequence is simple: the web and mobile are sufficient for tasks that reside in your connected tools and account files; the desktop remains the reference workstation as soon as the work involves documents stored on your disk.

The three steps to get started

Anthropic summarizes the handling in three steps. First step: open Claude, on claude.ai or in the desktop application, the latter being recommended whenever the work involves files from your computer. If your account still displays both modes, select " Cowork » in the message area.

Second move: throwing /setup-claude in a new conversation (the command replaces the old one) /setup-coworkA guided wizard helps you choose your role, installs a plugin tailored to that role (i.e., a package of skills, connectors, and example prompts), and then prompts you to connect the tools that this plugin uses, such as Slack, Google Workspace, or Microsoft 365. This is the fastest way to avoid starting from scratch; once the plugin is installed, type / in the message area to see the starting prompts he brought.

Third step: giving to Cowork a place to work. This is the one everyone underestimates, and it's the most structuring. Select "Work in a folder" in the chat bar. Two options: one dossier from your computer, including Cowork reads the content and in which it saves its new files; or a Project, a workspace that retains its own files, instructions, and memory from one session to the next. We'll come back to this in detail later, because this is where the difference lies between a tool that starts from scratch and a tool that knows you.

Connect your tools: connectors, browser, applications

A connector allows Cowork to read and act directly within a tool, find a Slack thread, open a document in Drive, update a record in the CRM, without having to search for and paste the information. You manage them in Customize > Connectors (Customize, Connectors). Common connectors include Slack, Google Workspace, Microsoft 365, Salesforce, and any internal connectors your organization may have published. The real power comes when multiple connectors work together: comparing a meeting transcript stored in Drive with commitments noted in another tool, and uncovering what was missed, becomes a ten-minute task.

For websites that do not have a connector, two browsers are possible, and you choose the preferred one in Settings > Cowork Preferred browser. The integrated browser The desktop application (macOS, Windows, Linux beta) opens a side panel where you watch Claude navigate, read, click, and fill out forms; it runs in an isolated profile, doesn't see your login credentials unless you import them on first launch (from Chrome, Edge, or Firefox on macOS, from Firefox on Windows and Linux), and then remembers your logins on that computer. It asks for permission before interacting with a site the first time and blocks high-risk sites. Claude in ChromeThe extension, on the other hand, operates within your actual browser, using your already open sessions. In both cases, Anthropic strongly advises against handling sensitive information, especially financial accounts or medical data.

The final lever, reserved for the Pro and Max plans: the managing your applications (computer use). When neither a connector nor a browser is suitable, Claude can open an application on your computer, click and type as you would, taking screenshots to orient himself. He activates in Settings > General > Enable computer use (macOS 15 or later, Windows) and requests your permission on a case-by-case basis; you can maintain a list of blocked applications, some of which, like investment platforms, are blocked by default. It's slower than a connector, still in beta, and there's no sandbox between Claude and your applications: reserve it for applications that have no other way to access them.

In Team and Enterprise organizationsThe owners decide which connectors are available, whether the integrated browser is active (enabled by default on Team; on Enterprise, disabled until September 9, 2026, and enabled by default from September 10 onwards, unless otherwise specified), and whether the separately managed Chrome extension is also enabled. Application management is not offered with these plans.

Files, Projects and Memory: Where Cowork put away what he knows about you

There are now four ways to provide context to Coworkand they are not equal.

The attached file The conversation is uploaded to the session environment. This fixes a pitfall from the first version: a file dragged from "Downloads" is now readable simply because it's copied to the cloud. It won't be modified on your hard drive; the result is returned to you as a downloadable file.

The connected folder stays on your computer. Claude can only read and write to the folders you have explicitly authorized via the desktop application; this is where he finds your source files and places his deliverables, alongside the originals. To get started smoothly, create a dedicated folder, " Cowork "In your Documents folder, for example, and work within it while you get your bearings: you'll always know what Claude sees and what he can access. The case instructions, detailed below, are also located there."

The project is the organizational unit. A Project Cowork It brings together instructions (tone, formats, rules applied to all Project tasks), files, context (local folders, linked chat projects, reference URLs), and its own scheduled tasks. Most importantly, Each project has its own memory Claude retains what he learns while working on tasks within a Project and reuses it in subsequent Projects, without it affecting other Projects. You can start from scratch, import an existing Project from the chat (one by one), or create a Project from a folder on your computer; in the latter case, the Project remains on that computer and does not synchronize with your account, whereas Projects created in Cowork They follow you from one device to another. On Team and Enterprise, a Project can be shared for reading or editing.

The account's memoryFinally, there's the new feature for the summer. Since the memory overhaul in the summer of 2026, Claude retains what he's learned about you from one conversation to the next—your role, your format preferences, your current projects—and applies it in the chat as well as in... Coworkwith one condition: that the session runs in the cloud (local desktop application sessions do not have access). It is enabled by default on the Free, Pro, and Max plans, and disabled by default on Team and Enterprise plans, where an owner must open it. You read, correct, and delete it topic by topic in Settings > MemoryThis is also where you'll find the "Generate memory from chats" switch, the option to include sensitive topics, the pause function, and the reset button. Conversations in incognito mode do not write anything to this area.

The key phrase from the first version, "the memory of Cowork The statement "exists only in Projects" is therefore outdated. The recommendation, however, still stands. Create a project before starting the real work., because it is the only place where instructions, files, memory and scheduled tasks of the same subject are brought together, and because the account's memory, general by construction, will never replace a well-maintained Project context.

The real secret: build a workspace, not a prompt

If you were to take away only one idea from this article, let it be this: In the cat world, the reflex is "faster, faster, better result." CoworkThe prompt is no longer the key element; The workspace is. Cowork It doesn't just accept what you type in the message box: it works with everything it can read in the folder or project you open—your files, your rules, your context, your communication style, your previous drafts, your success stories. The day your workflow is documented in files is the day the results stop being generic and become "it's like this tool already knows how I work."

There's no need to build an entire digital open-plan office. Four folders are enough to get started:

MON-ESPACE-DE-TRAVAIL/
├── system/      → les règles de base (assez stables)
├── context/     → le contexte stable (assez stable)
├── projects/    → le travail en cours (change selon les projets)
└── outputs/     → les livrables produits

A structure small enough to remain healthy, flexible enough to grow. But a single file doesn't tell much of a story. Cowork What makes it useful are two context files. HOW-I-WORK.md ("How I work") brings together everything a smart new colleague would need to know on the first day: who you are, what your organization does, who it serves, where your projects stand, the concrete realities of your daily life. SYSTEM-RULES.md (The "system rules") establish the safeguards: what Claude must always do, what he must never do, your preferred formats, your tone requirements. Since Cowork Having global, project, and folder instructions, these two files have a natural outlet: stable rules migrate to instructions, while the living context remains in the files.

A methodological tip from advanced users: Do not write these memory files. You would be describing an idealized version of yourself, not the one who actually works. Generate them through interviews: ask Claude to interview you, one question at a time, about your work process, then let him write the two files based on your answers. The result is much more accurate, and the exercise takes twenty minutes.

Then apply the principle of growth: start small, let the work dictate the rest. Only add a file or rule when a pattern repeats. If Cowork He systematically confuses a summary with a recommendation; in this case, an explicit rule is needed. Don't create files just because they might be useful someday. And delegate the writing: when a rule emerges during a session, simply say, "Codify this principle," and Claude will add it to the instructions for the file or project himself, without any intervention from you.

The three levels of continuing education

Beyond context files, Cowork offers three levels of instructions that apply without you having to repeat them. The overall instructions These apply to all your sessions. This is where you set your tone, output formats, working language, and a reminder of your role. You define these in Settings > Cowork, by clicking on "Edit" next to Global instructions, then "Save". Project Instructions In addition to all tasks of a given Project, there are: client writing rules, structure of deliverables for a project, in-house vocabulary. File instructions, available on the desktop application, accompany a local folder; Claude can update them himself over the course of a session, as he learns how you work.

While you're in the settings, go through Settings > Capabilities To check what's enabled (memory, web search, browsers), go to Settings > Memory to review what Claude has already learned from you. Five minutes that make a noticeable difference in all subsequent sessions.

Permissions and security: the chapter not to be skipped

Cowork It actually acts on your files, in your tools, on your behalf, and it reads content from outside. This calls for protections, and Anthropic has stacked several layers. The first is theisolation described above: a temporary cloud environment per session, without access to your networks, and access to your workstation limited to the folders you have opened, only through the desktop application.

The second is the permission modeThe most important daily setting, selected from a task-level switch, is the manual approval mode. There are now three options. "Manually approve": Claude pauses and asks for your approval before each action; this is the right choice for learning the tool, working on unfamiliar files, or closely monitoring progress. "Automatically approve": Claude proceeds, but a security check examines each action before execution and blocks those it deems dangerous; this is the safe operating mode for tasks you've mastered. "Skip all approvals": Claude doesn't stop, and nothing checks his actions; this should be reserved for low-stakes situations, on trusted files and sites, under your active supervision. In the unified experience deployed since September 16, these choices are labeled "Auto" and "Manual." Regardless of the mode, Claude always asks for your explicit permission before permanently deleting a file : the protection against deletion cannot be disabled.

The third layer targets theprompt injectionThe risk inherent to agents is that malicious instructions might be hidden within a web page, email, or document that Claude reads, attempting to hijack the task. The model is trained to distinguish your instructions from the content it processes; classifiers analyze untrusted content, and in automatic mode, a filter blocks suspicious actions. This risk only exists if two conditions are met: Claude reads content outside your trusted network, and it has the means to act accordingly. Hence, three key precautions: connect only the folders necessary for the task, be wary of unexpected access to a file or website, and do not use the browser or application management to manipulate sensitive information. Finally, note that the network egress settings you or your organization have defined do not apply to web searches, page retrieval, or MCP servers: the level of trust assigned to each MCP must be determined before installation.

For Team and Enterprise administratorsgovernance has been significantly strengthened since June. Cowork is enabled by default and can be disabled in the organization settings; the "Automatically approve" mode can be removed from the member selector; the "Always allow" option for connector tools is disabled by default; web search is disabled for the entire organization, chat and Cowork understood; skills can be analyzed for malicious code; plugins are distributed via an internal marketplace with four statuses (installed by default, available, mandatory, unavailable); events Cowork (Tool calls, file access, approval decisions) are sent to your SIEM via OpenTelemetry. And the major drawback of the first version is resolved: the sessions Cowork Launched from the web, the desktop and mobile applications are now captured by the compliance API.The only difference is that strictly local session data remains stored on users' machines, outside of standard retention policies, recoverable via this API but without remote deletion point.

The golden principle has not changed: Examine Claude's proposed action plan before letting him proceed, especially with sensitive files. Human oversight is not optional; it's the cornerstone of responsible use.

Plugins, skills, and scheduled tasks: from useful to transformative

Plug-ins adapt Cowork tailored to your role, your team, and your company: each package includes skills, connectors, sub-agents, and starting prompts. Installing one that matches your business means completing the setup with a real starting point rather than an empty session, and an organization can publish its own, customized to its tools and processes. The skills These are small instruction modules that activate themselves for a given task, such as producing a Word document that conforms to the charter, building a workbook with its formulas, or generating a presentation, and which transform the generalist into a specialist; Anthropic provides them, and you can create your own.

The planned tasks are, for many, the moment when Cowork becomes transformer, and this is the area that has changed the most. They now execute remotely, in the cloud, even when your computer is asleep or the application is closed The "computer on and application open" limit from June has disappeared. You create them from "Scheduled" in the sidebar, "New task", then "Create with Claude" (it asks a few questions and suggests a name, frequency, and instructions) or "Set up manually", or simply by typing /schedule in any task. Possible frequencies: hourly, daily, weekly, on weekdays, or on demand; you also choose the approval method and, optionally, the template. They have the same features as a regular task, including connectors, skills, and plugins, and are offered on all paid plans. A summary every morning at 7 a.m., a weekly report, recurring monitoring, team updates: all of this runs automatically.

Two rules of caution. A scheduled task works with your account's connectors and files, not with your computer's folders; if it needs them, it can only run when that computer's desktop application is open. And because no one is watching it run, start with low-stakes tasks, without sensitive data or actions with significant consequences, then review its past executions on the "Scheduled" page before expanding its scope.

A question of resources: Cowork consumes more

One last practical point, often discovered too late. A work session consumes more of your usage allowance than a conversation, and this makes sense: multiple steps, multiple file reads, sometimes multiple sub-agents running in parallel. Three best practices to avoid reaching your limits too quickly: group related tasks into a single session rather than creating multiple sessions; reserve the work session for tasks that justify file access and extended execution, and keep direct communication for everything else; monitor your usage in Settings > Usage.

To configure is already to adopt

All of the above describes how to configure CoworkLet us conclude by explaining why this matters. In our consulting work, we regularly see the same scenario: an organization opens an agent-based tool, employees type in a request, get a correct but generic result, and conclude that "it's not that different from a chatbot." They are right: without configuration, it isn't. And the merging of chat and... Cowork makes this scenario more likely, not less, since nothing signals to the user that they have just changed registers.

The lesson is that Configuration is not a preliminary technical chore: it is the first act of adoption. The moment a colleague takes thirty minutes to write down "how I work," set their safeguards, create their project, and choose their permission level is the moment they transition from passive user to active participant. The technical aspects can be mastered in a morning; what takes time, and what creates lasting value, is putting into words each person's way of working.

Your Getting Started Checklist, September 2026 version

1. Check your plan (Pro, Max, Team or Enterprise) and, on Team or Enterprise, that cloud sessions are enabled by your organization.
2. Install or update the desktop application (macOS or Windows) if your work involves local files; otherwise, claude.ai or the mobile application is sufficient.
3. Launch /setup-claude : role, plugin, connectors.
4. Create a Project for your first topic and write its instructions.
5. Create a dedicated folder on your computer, connect it, and build the minimal workspace: four folders, HOW-I-WORK.md et SYSTEM-RULES.md, ideally through interviews.
6. Set your global instructions in Settings > Cowork, then reread Settings > Memory and Settings > Capabilities.
7. Choose "Manually approve" for your first tasks; switch to "Automatically approve" when you know the terrain.
8. Configure your preferred browser and, if you have the use and the right, enable application control with an exclusion list.
9. Launch a first real task, but without stakes, in your dedicated folder, then a first planned task with low risk.
10. When a pattern repeats, add a rule. Not before.

Sources. Claude Help Center (accessed September 27, 2026): Get started with Claude Cowork, Use Claude Cowork on web, desktop, and mobile, Claude Cowork and chat are one Claude, Use Claude Cowork safely, Schedule recurring tasks in Claude Cowork, Organize your tasks with projects, Use Claude's chat search and memory, Use the built-in browser in Claude Cowork, Let Claude use your computer in Cowork, Use Claude Cowork on Team and Enterprise plans ; Claude Academy, Get started in Claude Cowork in three steps TechCrunch Anthropic merges Claude chat and Cowork in one interfaceSeptember 16, 2026. The workspace method (four folders, two files, interview) is that of our workshops.

For further : ten use cases of Claude Cowork by jobhow to choose between Claude, ChatGPT or Copilot and how build your personal assistant with Cowork.