How to set up Claude Cowork, step by step

Any paid plan. As of August 2026, Anthropic makes Claude Cowork available on Pro, Max, Team, and Enterprise, and not on the free tier. There is no separate Cowork subscription and no add-on charge, so the plan question is really a question about usage limits and administrative controls.

For a business, the practical floor is Team, at $20 per seat per month billed annually with a two-seat minimum, because Team is where central billing, single sign-on, role-based access, and the no-training default on your content begin. Pro at $17 per month billed annually is enough for one person evaluating the tool. The full plan arithmetic is in What Claude actually costs a small business.

One thing to size before you buy: Anthropic documents that complex Cowork tasks consume considerably more usage than ordinary chat. Two standard seats and one premium seat is a more realistic first configuration for a small office than three standard seats.

You usually do not have to. Cowork is on by default for Team and Enterprise organizations, and the administrative control turns it off, not on. What matters is knowing where that switch is before somebody asks you to use it.

1. 01Sign in as an Owner or Primary Owner of the Team or Enterprise organization.
2. 02Go to Organization settings, then Cowork.
3. 03Leave the switch on to keep Cowork available to everyone, or turn it off to disable it for every user in the organization at once.

If members report that Cowork is missing, this switch is the first thing to check, ahead of app versions and reinstalls. On an individual Pro or Max plan there is no organization layer, so this step does not apply.

Yes, for any work that touches local files, a browser, or the computer itself. Anthropic documents the desktop app as generally available on macOS and Windows, with the Windows build requiring the latest version, and offers ChromeOS and Linux downloads as well. The web version is in beta and rolling out to Pro plans, and mobile is in beta.

One detail is easy to miss. Cowork sessions run in the cloud, but the desktop app still has to be open and connected for Claude to reach local files, use a browser, or interact with the computer. Closing the laptop mid-task ends the parts of the job that depend on the machine.

Plan the rollout around the desktop app for daily work and treat web and mobile as ways to check on a job from elsewhere.

For any office with retention or audit obligations, this is the most consequential setting in the whole setup, and it is easy to pass over. Anthropic documents that a local Cowork session stores its conversation history on the user’s own computer, where it is not subject to standard data retention policies and cannot be centrally managed or exported by an administrator. Cloud sessions save projects with the member’s Claude account instead.

On Enterprise, cloud sessions are a deliberate grant rather than a default: an administrator turns on "Run Cowork in the cloud" and then grants the Cowork in the cloud capability to a group through custom roles. Enterprise organizations can also stream Cowork events to a SIEM through OpenTelemetry and retrieve remote sessions through the Compliance API.

Decide this before anyone starts, and write the decision into your AI use policy. A firm that owes a client an answer about where conversation history sits should not be discovering the answer afterwards. The section-by-section version of that document is An AI use policy for a small firm.

Set them from the Customize screen, under Connectors, or from the plus menu in the chat box, and set them before a task is running rather than while one waits on an approval. Each connector takes one of three modes.

Two administrative details belong in the same sitting. The organization setting that lets members choose "Always allow" for connector tools that can write is off by default, and leaving it off is the conservative starting position for an office still learning what these tasks do. Separately, Auto mode, which runs built-in safety reviews before actions execute, is a Pro and Max feature rather than an organization-wide one.

Administrators on Team and Enterprise can also stand up a plugin marketplace and set each plugin to Installed by Default, Available, Required, or Not Available. That is how a curated toolset stays curated once more than a handful of people are working this way.

Point it at a folder, and pick the folder deliberately. On desktop, Claude reads local files without manual uploading and can write back into the directories it has been given, which is the whole convenience and the whole risk in one sentence. A folder scoped to one matter, one client, or one project is a better first choice than the root of a document store.

Selecting a local folder makes folder instructions available, which are standing notes Claude reads every time it works there. Ten minutes here is worth writing into the plan, because folder instructions are where an office encodes its conventions once instead of restating them in every prompt.

• The naming convention for files written into the folder, so output lands where a person expects it.
• The house format for the deliverable: the sections a memo carries, the fields a summary always includes.
• What must never be touched or overwritten in that folder.
• Who reviews work produced there before it leaves the office.

Choose Cowork in the message box, describe the outcome you want rather than the keystrokes to get there, then read the approach Claude proposes before permitting it to run. Reviewing that plan is the step worth protecting once the tool stops feeling novel, because it catches a misread instruction while it is still cheap.

1. 01Pick a task with a checkable answer for the first run: something where you already know what good looks like.
2. 02Describe the outcome, the source folder, and the format you want back.
3. 03Read the proposed approach. Correct the scope before permitting execution, not after.
4. 04Watch the steps as the task runs, since each one is shown, and stop it if the approach drifts.
5. 05Check the output against the source yourself. The first task is a calibration exercise, not a delivery.

One behavior to know before it surprises somebody: memory does not carry between ordinary chat and Cowork, except within Projects, and sessions cannot currently be shared with a colleague. Context that matters belongs in folder instructions or in a Project rather than in a conversation somebody else cannot open.

For which tasks are worth handing over first, What small teams actually do with Claude Cowork works through the catalog by role.

Four decisions, and all four were made during the setup above, so writing them down costs a page rather than an afternoon. Doing it now is what keeps the fifth person to get a seat working the same way as the first.

1. 01Which folders and connected systems Cowork may reach, and which are out of scope.
2. 02Whether local sessions are permitted, given that their history sits on the user’s own computer outside standard retention.
3. 03Which connectors are set to needs-approval, and who may change that.
4. 04Who reviews Cowork output before it leaves the office, for each category of work product.

Justo does this setup as part of a flat-fee engagement: plan selection, secure setup with access scoped deliberately, training on your real work, and the signed AI use policy, from $3,500 for a team of 5 to 25. Scope is on the pricing section of our main page.

Four, all documented by Anthropic.

None of these are reasons to wait. They are reasons to choose the first task, the first folder, and the first seat mix on purpose. What Cowork is and how it differs from ordinary chat is covered in Claude Cowork, explained.