◈AI Codex
Administrationupdate

Computer use comes to Team and Enterprise: what admins control before turning it on

In brief

On November 20, 2026 Anthropic put computer use into beta on Team and Enterprise plans. Claude can see a member's screen and click and type in the desktop apps they approve. It is off by default and only an Owner can turn it on. The org toggle, role exclusions, app block and allow lists, the MDM switch, what none of them cover, and a rollout order for the first two weeks.

9 min read··Computer Use

Contents

♡Sign in to save

On November 20, 2026, Anthropic put computer use into beta on Claude Team and Enterprise plans. It is rolling out to organizations gradually, so the setting may not be in your admin panel yet.

Computer use lets Claude see a person's screen and work in the desktop apps that person allows. Claude takes screenshots, moves the pointer, clicks, and types, the way the person would. It exists for the work that has no connector: the legacy ERP client, the vendor portal installed as a desktop app, the internal tool nobody will ever build an integration for.

It is off by default. This article is for the person who decides whether to turn it on.

Where it works

  • Claude Desktop on macOS 14 or later, and on Windows.
  • Claude on the web and Claude Mobile, when your organization's "Computer access" setting is on and the chat is connected to a computer that has Claude Desktop open. The work still happens on that computer.
  • Only in the new Claude experience, so "Chat and Cowork unified" has to be on for your organization first. If you have not made that switch, read Cowork and chat are now one Claude.

It does not work in tasks started from the separate Cowork option, or in Claude Code.

Some organizations cannot turn it on at all: those with a HIPAA-ready configuration, Zero Data Retention for Claude Code, customer-managed encryption keys, or FERPA or K-12 settings.

The three controls you have

All three live in Organization settings > Capabilities, under "Claude Desktop". Only an Owner or Primary Owner can change them.

1. The organization toggle

"Computer use" is one switch, off by default. Turning it on is fast. Turning it off can take about an hour, sometimes longer, to reach every member's Claude Desktop. Plan for that delay if you ever need to pull it in a hurry (the MDM switch below is the faster stop for a single machine).

2. Role exclusions (Enterprise only)

On Enterprise, computer use is a capability inside custom roles. When you turn it on, a list of roles opens and you clear the ones that should not have it.

Two details catch people:

  • The built-in roles (User, Admin, Owner, Primary Owner) get computer use the moment it is on. To exclude one person, move them to a custom role and wait 15 minutes.
  • A role set to "All capabilities" cannot be cleared from the list. Change its "Capability access" to "Only selected", turn off "Computer use" on that role, wait 15 minutes, then turn the feature on for the organization.

Team plans have no custom roles. On Team the switch covers everyone.

3. App access

Under "Computer use" there are two lists. Only one is active at a time.

List Effect Use it when
Block these apps Listed apps are denied. Everything else is available. You trust the feature broadly and want to fence off a few sensitive apps.
Allow only these apps Listed apps are available. Everything else is denied. You are piloting on one or two named workflows.

Either way, each member still approves each app the first time Claude asks to use it in a session.

You add apps by searching, or by typing IDs one per line: the bundle ID on Mac (com.microsoft.Excel), the file name on Windows (excel.exe) or the Microsoft Store ID. Nothing checks what you type, so a typo fails silently. Test every entry on a real machine.

Three things to know before you rely on a list:

  • Changes take about an hour to reach members.
  • With an allow list active, anyone not on the latest Claude Desktop loses computer use until they update.
  • The list is sent to every member's computer and can be read there. Do not put a confidential app name in it.

The device-level switch

If you manage laptops with an MDM tool (Jamf, Intune, and similar), you can turn computer use off on a specific machine regardless of the organization setting. Set the policy key computerUseEnabled to false. It needs Claude Desktop 5.1231.0 or later and takes effect after Claude Desktop is quit and reopened.

MDM can only turn it off. It cannot turn computer use on, approve apps for a user, or grant screen recording.

On Mac, each member has to grant Claude Desktop two system permissions: Accessibility, and Screen & System Audio Recording. Non-admin users may need an admin password to do that. If your Mac privacy profile sets Deny on either one, computer use breaks without an error message. Check your profile before the pilot starts, or your first support ticket will be "it does nothing."

Windows needs no system permission.

What the controls do not cover

Anthropic's setup page is direct about this, and it is the part to read twice.

  • There is no sandbox. Nothing sits between Claude and the apps it is allowed to use. Text on screen can steer Claude into an action nobody asked for. That is prompt injection, and it is the main risk of the feature.
  • Folder limits do not apply. A folder marked "Read only" for Claude's file access is not read-only to computer use.
  • Deletion protection does not apply.
  • App lists stop at the app boundary. They do not cover Claude in Chrome, the built-in browser, Claude Code, desktop extensions, websites, apps inside a remote desktop or virtual machine, or a member's personal Claude account on the same laptop. A link clicked in an allowed app can open in another app Claude has no permission for.
  • Screenshots are saved with the chat. They can contain whatever was on screen: a customer record, a salary figure, a private message. Anthropic's automated action check also sees the latest screenshot and can keep a copy.

Anthropic's own advice is to keep computer use away from financial accounts, legal documents, medical information, and apps holding other people's personal data.

The help pages do not state how long screenshots are kept, whether computer use actions appear in the Compliance API, or whether the feature changes usage limits. If any of those is a requirement for you, ask your Anthropic account team before the pilot, and get the answer in writing.

What Claude will and will not do on its own

Claude tries a connector first, then a browser, and uses the screen last, because screen work is the slowest of the three.

Inside allowed apps, the defaults are narrower than "full control":

  • Browsers and many trading and crypto apps: Claude can look but cannot click or type.
  • Terminals and code editors: Claude can click and scroll but cannot type. A click can still run something.
  • Music, video, and e-book apps are off limits.
  • Claude is trained to avoid typing passwords and card numbers and to check with the person before sending a message or submitting a form. Anthropic says these safeguards are not perfect.

The computer has to be awake with Claude Desktop open. Long multi-step tasks sometimes need a retry.

A rollout order

Before you touch the toggle

  1. Pick one workflow with no connector and low blast radius. Copying figures from a desktop reporting tool into a spreadsheet is a good first case. Anything that sends, pays, or deletes is a bad one.
  2. Write down which apps that workflow needs. That is your allow list.
  3. Update your AI usage policy with two lines: which apps computer use is approved for, and that members close anything sensitive before starting a task, because whatever is on screen ends up in a screenshot.

Week one

  1. On Enterprise, create a custom role for the pilot group so that only those five to ten people have the capability.
  2. Turn on "Computer use" with "Allow only these apps" active and your short list in it.
  3. Sit with two pilot users for their first run. Watch the Mac permission step and the per-app approval. Those are where people get stuck.

Week two

  1. Ask each pilot user one question: which task did you hand over, and did you have to redo it? A task that needs a retry more than about one time in three is not ready.
  2. Decide whether to widen the allow list or switch to a block list. Switch to a block list only once you have a named list of sensitive apps to put in it: your finance system, HR system, password manager, and anything holding customer personal data.

If you are on Team, you have no role exclusions, so step 4 is not available. Use the allow list as your pilot boundary and tell the wider team in plain words that the feature is on but approved for one workflow.

For the broader pattern of running a first test with a small group, see Running your first AI pilot. For the other Desktop and security settings that shipped this month, see the Enterprise security update.

Sources: Claude release notes, November 20, 2026, Set up computer use for Team and Enterprise plans, and Use Claude Cowork safely.

Related tools

Weekly brief

For people actually using Claude at work.

Each week: one thing Claude can do in your work that most people haven't figured out yet — plus the failure modes to avoid. No tutorials. No hype.

No spam. Unsubscribe anytime.

What to read next

All articles →
Computer Use·update·11 min

Computer use goes GA, and a new browser tool ships alongside it

On August 19, 2026, Anthropic moved computer use to general availability as computer_toolset_20260801 and launched browser use as a separate client toolset. The Skills API and Files API went GA the same day. Here is what changed in the request shape, when to pick browser over computer, and what to check before you migrate.

Computer Use·How It Works·11 min

Claude in Chrome vs. the Cowork browser: two ways to let Claude use the web, shipped the same day

On August 26, 2026 Anthropic made Claude in Chrome generally available on every paid plan and gave Cowork its own built-in browser. They look like the same feature and are not. Chrome works on the pages you already have open, with your logins, and now auto-approves actions a classifier judges safe. The Cowork browser is a separate, isolated browser Claude drives alone. Here is which one to use for what, and where the prompt injection risk actually sits.

update·9 min

Claude Enterprise security update: a security center, parent organizations, and sign-out everywhere

On November 5, 2026 Anthropic shipped seven security and admin features for Claude Enterprise: a security center that checks your settings against Anthropic's recommendations, centralized management for companies with several Claude organizations, an Admin API call that signs a member out of every device, a self-serve IP allowlist, Desktop settings and outside-account blocking, inference hooks out of beta, and EDR inside Cowork. What each one does, who can turn it on, and the order to do it in.