On September 16, 2026, Anthropic announced that Claude Chat and Cowork are merging. The rollout starts with Pro and Max over the following weeks. Work that previously required choosing Cowork becomes available from an ordinary conversation, using the context and connections the account already has.
Start with the permissions already on the account. A folder or connector approved for an earlier task may still be available when someone starts a new conversation. Review that access and any unattended work before handing over another task. This procedure draws on Anthropic's documentation; General Analysis has not tested the migration.
Establish which experience the account has#
Anthropic's migration guide says Pro/Max accounts move in stages. If the message box still offers separate Chat and Cowork choices, that account has not moved. After migration, users cannot switch back. Existing tasks, connectors, skills, and projects carry over.
Record the plan, account, client version, and visible experience before comparing settings. Two users on the same plan may be in different rollout cohorts. A screenshot from a migrated personal account cannot establish what an Enterprise member sees.
The September announcement promised Enterprise administrators advance notice. By our October 19 review, Anthropic's Enterprise setup guide documents an opt-in public beta named Chat and Cowork unified (beta). It is off by default. Record whether the organization or a custom role enables it; the original announcement does not establish a tenant's activation date.
Anthropic requires Chat plus Cloud code execution and file creation for this beta. The separate Cowork and Cowork in the cloud capabilities are not prerequisites. Organizations using ZDR, the HIPAA-ready configuration, or customer-managed encryption keys are currently ineligible. Keep those organizations on their existing configuration rather than treating the consumer rollout as authorization to migrate them.
Detect and respond to employee AI agent risk
General Analysis AI Detection and Response monitors employee AI agents across files, browsers, MCP servers, connectors, scheduled work, and downstream business actions. It gives security teams the trace, policy decision, and response workflow needed to stop unsafe data movement.
Separate connector loading from action approval#
Two settings named Auto can lead to a mistaken access review. Tool access controls when connectors enter the conversation. The permission mode controls how actions receive approval.
| Setting | What it decides | What to review |
|---|---|---|
| Tool access: Auto | Claude chooses which connectors to load | Which connected services remain eligible |
| Tool access: Always available | Connectors load at conversation start | Whether every connection still has a purpose |
| Tool access: On demand | Claude searches for relevant connectors before loading them | Do not treat delayed loading as revoked access |
| Permission mode: Manual | Actions follow manual approval rules | Saved tool preferences and the actual prompt behavior |
| Permission mode: Auto | Automated safety checks evaluate actions | Whether unattended action authority fits this task |
The consumer migration guide describes Manual as the default, with the permission setting applying to the whole conversation. Anthropic's Cowork permission matrix qualifies that behavior: in Manual mode, a connector tool marked Always allow is approved, while a tool marked Needs approval asks. Blocked tools remain denied. For an Enterprise pilot, check the organization controls below as well.
Choose Manual for the initial review, then inspect the tool preferences underneath it. Merely seeing “Manual” is insufficient evidence that every consequential connector action will produce a fresh prompt. Likewise, changing tool loading to On demand leaves the connection available for Claude to discover.
Consider a sample task: summarize a project update and leave the draft in the conversation. Reading the connected document and posting the draft to a shared channel require different authority. Check whether the account can do both before using real project material. Asking for a summary should not authorize its distribution.
Review what carried over#
Use the migration guide's new settings locations to inspect the account:
- Open Customize > Connectors and identify the services still connected. For each, record the connected identity and business purpose, then remove access that is no longer needed.
- In Desktop settings, review Trusted folders. Give each retained folder an owner and a reason to remain accessible; replace broad grants with a dedicated working folder where practical.
- Check Instructions for Claude under Settings > General. Cowork's Global instructions move there. Remove obsolete workflow assumptions, but enforce sensitive access through permissions and the connected system.
Review browser and computer-use permissions separately from connector tools. Anthropic's safety guidance warns that computer use acts on the screen without the same permission checks that gate other tools. A connector's approval preference therefore does not establish what a permitted desktop app can do.
For an organization's account-boundary cleanup, use the separate Claude account restrictions guide. For device, browser, and network controls, use how to secure Claude Cowork.
Pause unattended work explicitly#
Closing Desktop cuts off access to the computer while it is offline. It does not establish that cloud work has stopped. Anthropic documents cloud sessions and server-side processing, including processing of local files opened through the Desktop bridge. Taking the device offline cannot recall data already read.
Open Scheduled and review each task's instructions, cadence, approval mode, and recent runs. Scheduled tasks have access to configured connectors and plugins. Pause or delete unused schedules through those controls. Inspect currently running sessions separately; pausing a schedule should not be recorded as proof that an already-started run ended.
The scheduling page says remote tasks cannot be tied to local folders, but its manual setup section also describes local folders. Do not assume every inherited task has identical execution requirements. Record whether the specific task uses cloud files, connectors, or a local dependency, and leave that field unresolved if the interface and documentation disagree.
For the first review, keep schedules limited to harmless sample inputs and draft outputs. Anthropic recommends avoiding unattended work involving sensitive information or consequential actions. Review the next run before restoring a useful schedule to normal operation.
Prepare Enterprise acceptance without assuming policy parity#
Use the unified Enterprise setup guide for the beta's controls. The older Team and Enterprise Cowork page describes the separate experience; do not assume identical approval behavior from similar setting names.
Anthropic documents these boundaries for the unified beta:
| Control | What to verify before a pilot |
|---|---|
| Organization or role capability | Confirm who receives the unified experience. The Admin API identifies the role permission as chat_cowork_unified; access is checked on every request. |
| Network egress and connector policies | Code runs in an isolated cloud VM. Egress settings govern that code; Anthropic enforces connector policies outside the sandbox on every connector call. |
| Connector standing approval | When Always allow for connector tools is off, members cannot grant standing approval and Claude asks each time. Inspect the actual policy and prompt. |
| Automatic action approval | Review Automatically approve and Skip all approvals separately. Anthropic specifically applies the latter to scheduled tasks and the built-in browser. |
| Sharing | A conversation that used its sandbox requires both Share chats and Share cloud sessions to be enabled. |
| Local access and schedules | Linked-computer access is optional. Computer access and Scheduled tasks controls are organization-wide, not per-role; a role-limited beta pilot does not narrow those settings. |
Before enabling a pilot, verify one permitted draft-only workflow and one harmless write that requires approval. Save the effective role, connector preference, and organization policy with the result. These are proposed acceptance checks; we have not tested Anthropic's enforcement.
Retention needs a separate decision. Anthropic says cloud sessions started within a conversation are deleted with that conversation under its retention practices. The sandbox expires after 30 days, but that does not delete the conversation. Standalone cloud sessions and scheduled task runs fall outside custom conversation retention. During the public beta, memory also stops following the organization's custom retention setting. Review those exceptions before allowing sensitive work, and verify exports through the Compliance API, which Anthropic says covers unified conversations using existing access keys.
Record the review outcome#
Keep a short record for each account or pilot group. The example below is a review format, not a Claude settings file. Fill its unresolved fields from observed settings and authorized checks.
Code source: illustrative.
review_id: chat-cowork-migration-01
account_plan: Pro
experience: unverified
reviewed_at: null
action_approval_mode: unverified
connector_loading_mode: unverified
retained_access:
connectors: []
trusted_folders: []
browser_and_app_grants: []
scheduled_work:
inventory_complete: false
unnecessary_schedules_paused: false
active_runs_reviewed: false
sample_workflow:
expected_output: conversation-draft-only
observed_result: unverified
evidence_reference: null
review_owner: null
Add the client version and relevant policy settings to the evidence reference. Keep secrets and private document contents out of the record. An empty connector list means the inventory is unfinished until the reviewer confirms otherwise.
Close the review when retained access has a purpose, unnecessary schedules are paused, and the permitted workflow behaves as expected. Assign an owner to any unresolved grant or policy difference before widening the rollout.
