Claude Cowork for business teams is on by default, and only an organization owner can switch it off, for everyone at once[1]. The setting to handle before you let it run is per-task approval: “Always allow” for connector tools is off by default[1]. Then decide what you can review afterwards[1].

This is for the operations manager or COO who has to answer for what the assistant touched. The question here is not what Cowork can do for a team. It is what you fix as an admin, and what the documentation says you can see once people start working. The builder’s tool is covered in Claude Code for business, and the file format behind a skill in Claude skills for business.

What is Claude Cowork?

Claude Cowork is Anthropic’s product, generally available on all paid plans, and Anthropic reports that most usage comes from outside engineering: operations, marketing, finance and legal[2]. An admin manages it per organization, through one switch in the organization settings[1]. One paragraph, on purpose. The rest of this page is about those settings.

The admin checklist before you switch Cowork on

Eight settings, with the default the documentation gives. The last column is yours; this table says what the product does, not what your sector permits.

SettingWhere you set itDefaultWhat you decide
Cowork for the organizationOrganization settings > Cowork[1]On, and off in organizations that enabled HIPAA[1]Whether every member gets it, or nobody
Groups and custom rolesAdmin console, group membership manual or through SCIM[2]No group scoping in placeWhether one team starts first, which needs an Enterprise plan[1]
“Always allow” for connector toolsOrganization setting, applies to members[1]Off[1]Whether members may skip approval for connector tools that can write
Permissions per MCP connectorAdmin console, applied org-wide[2]Set per connectorWhich system the assistant may change, and which it may only read
Network egressThe network settings your organization already has[1]Inherited, and a change applies to new sessions only[1]Nothing extra here, but note that web fetch, web search and MCPs fall outside it[1]
PluginsThe same admin toggle as Cowork[1]Follows Cowork, no separate setting[1]Nothing separately, so weigh plugins when you weigh Cowork
Group spend limitsAdmin console, per team[2]No limit setWhich team gets which budget
What you can review afterwardsCompliance API, admin dashboard, Analytics API, OpenTelemetry[1][2]Sessions via Claude, Claude Desktop and Claude Mobile are captured[1]Where sessions may run, since local history stays on the user’s own computer[1]

Rather walk that table against your own stack with someone? WeAdapt does that in a first session. Book a call.

Is Cowork on by default, and who can turn it off?

On by default, and only an organization owner can disable it. The admin toggle sits in Organization settings > Cowork[1], and it covers the whole organization: switch it off and no user in the organization has Cowork[1]. In organizations that have enabled HIPAA, Cowork is off by default[1].

That all-or-nothing shape decides the order of your rollout. On Enterprise plans there is a way around it: groups and custom roles let an admin enable Cowork for specific teams[1]. Admins on Claude Enterprise can organize users into groups, manually or through SCIM, and record per group which capabilities that group may use[2]. Without that plan the choice is the whole organization or nobody, which is worth knowing before you promise a pilot with one department.

Plugins travel with the switch. They are included with Cowork and controlled by the same admin toggle, and there is no separate setting for them[1].

Whether a team can carry the agreement that comes with the switch is a people question, not a settings question, and that is what a training or workshop is for.

Can a member write without approval?

Only if an owner turns on “Always allow” for connector tools. That organization setting is off by default, so a task using a connector tool that can write waits for the member to approve it[1]. Switch it on and members may skip that per-task approval for those tools[1].

The exemption is narrower than it reads. A read-only tool is exempt only when the connector annotates them as read-only[1]. And most custom connectors don’t annotate their tools, so every tool on those connectors is gated[1].

Read that twice if your team built its own connector. An MCP connector, MCP being the Model Context Protocol, that nobody annotated puts every call behind the approval prompt, including the harmless read that happens all day long. The fix sits in the connector rather than in the setting: annotate the read-only tools, so the gate falls where it belongs. How that build goes is in connect AI to your own systems, and WeAdapt builds the annotated version as a custom AI tool.

What do you set per connector?

Which actions the connector may perform. An admin can limit the allowed actions per MCP connector, for example allowing reads and blocking writes; those connector permissions apply org-wide and are configured from the admin console[2].

ActionWhat you setWhat to check first
ReadingAllow read actions per connector in the admin console[2]Which systems hold data you would rather keep out of a session
WritingBlock write actions on the systems where a change is not acceptable[2]Who would notice a wrong change, and how fast
ScopeNothing: permissions apply org-wide[2]Whether one team needs an exception, which then runs through groups[1]
Outbound trafficNothing new: Cowork follows your existing network egress permissions[1]That web fetch, web search and MCPs fall outside those permissions[1]

Outbound traffic is the part that gets skipped. A change to those network settings applies to new sessions and leaves running conversations alone[1]. So the connector permissions, not the network settings, are what sit between the assistant and a system you want left untouched. Which systems belong in the first round is the question WeAdapt works through in consultancy.

What can an admin see afterwards?

Sessions through Claude, Claude Desktop and Claude Mobile, which are captured in the Compliance API[1]. Local sessions are the gap: Cowork stores that conversation history on users’ computers, it falls outside Anthropic’s standard data retention policies, and admins cannot centrally manage or delete it[1].

Where the session runsWhat is recordedWhere you read it
Claude, Claude Desktop, Claude MobileThe session itself[1]Compliance API[1]
Local sessionsConversation history on the user’s own computer, outside standard retention[1]Nowhere centrally, and no admin can delete it[1]
Team and Enterprise plansTool calls, file use, skills used, and whether an action was approved manually or automatically[2]An OpenTelemetry stream to your SIEM, for example Splunk or Cribl[2]
Usage and spendActivity per user, use of skills and connectors[2]Admin dashboard and the Analytics API[2]

Two things about that table. OpenTelemetry does not replace audit logging for compliance purposes, which the documentation states plainly[1]. And the row with skills in it is the one that tells you what your team actually leans on, which is the same signal you use when you roll out a skill in the first place[2].

One warning about reading up on this topic. Secondary write-ups of the same feature claim the opposite on exactly this point, that Cowork activity does not land in the Compliance API. The vendor documentation says it does, for sessions through Claude, Claude Desktop and Claude Mobile[1]. Vendor documentation wins, and the date belongs to the fact: these settings were read on 1 November 2026, and a product this young moves.

Frequently asked questions

Can we switch Cowork on for one team only? On an Enterprise plan, yes: groups and custom roles let an admin enable Cowork for specific teams[1], and group membership is handled manually or through SCIM[2]. On the other paid plans the toggle is organization-wide, so it is all members or none[1].

Do plugins need a separate decision? No. Plugins are included with Cowork and controlled by the same admin toggle, and there is no separate setting for them[1]. So the decision about Cowork is the decision about plugins, which is worth saying out loud before anyone asks for one.

Can an admin delete what someone did in a local session? No. For local sessions Cowork stores the conversation history on the user’s own computer, outside Anthropic’s standard data retention policies, and admins cannot centrally manage or delete it[1]. Sessions through Claude, Claude Desktop and Claude Mobile are captured in the Compliance API[1].

A workable order: decide whether Cowork goes on for the organization or for one group, leave “Always allow” off, set the connector permissions before the first task runs, and agree where sessions may run. Four decisions, and only the local history is hard to undo.

Want the admin side settled before your team starts? WeAdapt reads the settings against your stack, sets them with you, and writes down who decides what. Book a call.

Sources

  1. Anthropic, 13455879 use claude cowork on team and enterprise plans (support.claude.com, accessed 1 November 2026)
  2. Anthropic, Cowork for enterprise (claude.com, accessed 1 November 2026)