Claude Cowork is the part of Claude that works through a multi-step task on its own: it reads the files and connectors you give it, runs code in a sandbox, produces documents and spreadsheets, and reports back, using the same agentic architecture as Claude Code without a terminal. Setting it up for a business means four decisions before anyone starts a task: where sessions run (cloud by default), which folders and connectors Claude may reach, which approval mode staff use, and what the admin console locks for everyone. This checklist walks through each one using Anthropic's documentation as of October 2026.
If you are deciding between Cowork and Claude Code for a team, read Claude Cowork vs Claude Code first. If your question is whether Cowork is safe for company files at all, that answer has its own post. This one is about doing the setup properly; the commercial side of a rollout lives on our Claude consulting and implementation service.
What is Claude Cowork, and who gets it?
Anthropic's Get started with Claude Cowork article describes it as bringing Claude Code's agentic capabilities to knowledge work beyond coding. Chat answers one prompt at a time; Cowork keeps going through a task, coordinates sub-agents, and can produce Excel files with working formulas and PowerPoint decks. The same article says that for Pro and Max subscribers Cowork is now simply Claude, with the two merged into one experience, and that the merge is in beta for Enterprise behind an admin setting called Chat and Cowork unified. Team organizations, and Enterprise organizations that have not turned the setting on, still see chat and Cowork as separate modes.
Cowork is on paid plans only: Pro, Max, Team and Enterprise. There is no separate Cowork price; it comes with the plan. Where it runs, as of October 2026:
- Claude Desktop for macOS and Windows on all paid plans. Windows needs the latest version of the app. The support article does not list a Linux desktop build for Cowork.
- The web at claude.ai on Pro, Max and Team, and on Enterprise where an admin enables it.
- iOS and Android on Pro, Max and Team, and on Enterprise where enabled.
- The Claude in Chrome side panel on Max and Team, rolling out to Pro, and on Enterprise where enabled.
So Cowork is available for Windows, through the desktop app, the web and mobile rather than a separate download, on any paid plan.
Step 1: Decide where sessions run (cloud versus local)
Anthropic's Cowork architecture overview describes two execution models:
| Cloud sessions (the default) | Local sessions (legacy) | |
|---|---|---|
| Where the work happens | On Anthropic's servers. Each session gets its own temporary sandbox, created at the start and destroyed at the end; sandboxes do not share state across sessions or organizations | The agent loop runs on the employee's computer; code runs in an isolated Linux virtual machine (Apple Virtualization.framework on macOS, Hyper-V on Windows) |
| Who has it | Pro and Max; Team in beta on web and mobile; Enterprise once an owner enables it | Pro and Max only, and only for tasks and scheduled tasks started on the device before October 6, 2026 |
| Reaching the employee's files | Through the Claude Desktop app over an Anthropic-brokered connection, limited to connected folders, with each local tool call checked against the member's permissions. If the desktop app is offline, the session cannot reach the device | Direct, within connected folders |
| Network | All outbound traffic passes through a mandatory proxy; only allow-listed destinations are reachable; private, internal, link-local and cloud-metadata addresses are blocked. No network access is the default for Enterprise organizations | The VM has its own egress filtering and syscall restrictions |
| Credentials in the sandbox | Session-scoped tokens that expire within hours; connector authorization tokens never enter the sandbox, because connector calls are made server-side | Not described in the same detail |
| What your endpoint tools see | Nothing: EDR cannot observe cloud sessions | Nothing inside the VM: EDR cannot inspect activity there either |
For a business the practical reading is: plan for cloud sessions, because that is where new work runs, and treat the admin network policy and the Compliance API, not your endpoint agent, as the controls. Team and Enterprise admins can turn cloud sessions on or off separately from local desktop Cowork at Organization settings, under Cowork.
Step 2: Connected folders, connectors and plugins
On the desktop app, Claude reads and writes local files only inside folders you connect. Start each pilot user with one working folder for the task, not their Documents root, and use Folder instructions (a note that gives Claude context for that folder) to tell it what the files are and what it may not touch. Claude requires explicit permission before permanently deleting files.
Connectors are the bigger decision. On Team and Enterprise, only owners can enable connectors such as Google Drive, Gmail, GitHub, Slack and Microsoft 365, at Organization settings under Connectors; members then connect their own accounts. Custom connectors (remote MCP servers) follow the same rule: on Team only Owners and Primary Owners can add one, and Anthropic's custom connector article warns that these reach services it has not verified, that a malicious server may carry prompt injections, and that during Research Claude can call connector tools without further approval. Before enabling one, write down who hosts it, which scopes it requests, whether its tools can write or delete, and whether any fixed credential is shared by everyone who uses it. We cover that review in how to review a Claude MCP connector, and the Microsoft 365 connector's permission list in Claude Microsoft 365 connectors.
Plugins bundle skills, connectors and sub-agents, and work in chat, Cowork and Claude Code. On Team and Enterprise, owners provision skills for the whole organization at Organization settings, under Plugins and skills, and can turn off user-created skills on the Policy tab.
Step 3: Approvals, network egress and device policy
Cowork has three approval modes. Manual, the default in the current experience, asks before Claude acts. Auto approves read-only tools and runs a safety review on other actions, blocking the ones it judges unsafe; it uses more of the plan's usage. Skip, from the older experience, asks nothing. Start every pilot in Manual. For the organization as a whole, Team and Enterprise admins can turn off the persistent Always allow option, so each permission-gated call needs a fresh approval, and can remove the Auto option entirely.
Network egress for cloud sessions is set by the admin: which destinations the sandbox may reach, or none. Two things to know before you write that policy. First, Enterprise starts at no network access, so a task that needs to fetch a vendor's price list will fail until you allow that host. Second, the egress setting does not apply to web fetch, web search, MCP connectors or Claude in Chrome, each of which owners can disable separately. A sandbox allowlist is not a web-browsing block.
Device policy is where IT comes in. Admins can require trusted-device enrollment and a recent sign-in for all cloud sessions. On managed computers, two MDM keys for the Claude Desktop app matter: isLocalDevMcpEnabled set to false disables plugin-bundled and locally configured MCP servers, and isDesktopExtensionEnabled set to false blocks MCPB and DXT extension servers. These are device settings, not organization settings, so they go out through Intune or Jamf with the rest of your baseline; local MCP servers do not run in cloud sessions.
Step 4: Pilot and handoff checklist
- Pick the plan and the workspace. Cowork on Team gives you SSO, connector controls and spend caps; Enterprise adds audit logs, the Compliance API, custom data retention and role-based access. Claude Team vs Enterprise has the comparison.
- Write the folder and connector list. One folder per pilot task, connectors read-only where the product allows it, custom connectors reviewed and approved by an owner.
- Set the organization controls. Cowork toggle, cloud sessions on or off, network policy, Always allow and Auto disabled for the pilot, trusted-device enrollment on, web search and Chrome decided deliberately.
- Push device policy. The two MDM keys above, and the latest desktop app version on Windows.
- Choose five to ten pilot users and three tasks. Recurring, document-heavy work with a clear finished state: a monthly report pack, a contract summary set, a data clean-up in a spreadsheet. Give each task a success measure before it starts.
- Run for two to four weeks in Manual mode. Keep a log of approvals people found pointless and approvals that stopped something wrong; that log becomes your allowlist.
- Review the evidence. Cowork activity is captured in the Compliance API, an Enterprise feature; Anthropic's architecture overview says Team and Enterprise admins can use OpenTelemetry. Check who used which connector and what left the sandbox.
- Hand off. A one-page standard: approved folders and connectors, approval mode, who can add a connector, how to request a skill, how to report a bad result. Then set up the first scheduled tasks, which run in the cloud without the device online.
What NetSys does for a Claude Cowork rollout
NetSys has run a Microsoft Copilot and Claude rollout for a 25-employee company with usage policies, connector access controls and staff training, and we set a lean team up on Claude with data-handling rules in place before building custom apps and reporting roundups on it. For Cowork, our Claude consulting service covers the workspace choice, the organization and device settings above, the connector review, the pilot and the handoff document; Claude training covers the people side. Where you want the controls checked independently before staff use Cowork on company files, the AI security assessment reviews what Claude can reach, what it may do and what the logs will show, within a scope agreed first.
Request a Claude rollout review and bring the tasks you want Cowork to take on; we will tell you which settings they need.
Frequently asked questions
How do I access Claude Cowork?
Sign in on a paid plan. On Pro and Max it is simply part of Claude; on Team, and on Enterprise where the admin has enabled it, you choose Cowork in the desktop app, on the web or on mobile. There is nothing separate to install beyond the Claude Desktop app, which is where connected folders live.
Is Claude Cowork available for Windows and Linux?
Windows, yes: the Claude Desktop app for Windows runs Cowork on all paid plans and needs the latest version. Anthropic's Cowork article lists macOS and Windows for desktop, plus web, iOS and Android; it does not list Linux as of October 2026.
How much is Claude Cowork?
It has no price of its own. As of October 2026 Anthropic's pricing page includes it in Pro ($17 a month billed annually, $20 monthly), Max (from $100 a month), Team ($20 per standard seat a month billed annually, $25 monthly) and Enterprise ($20 per seat a month billed annually, with usage billed at API rates). Our Claude pricing for a business post works through the seat math.
What can Claude Cowork do that chat cannot?
Carry a task through several steps without a prompt for each one: read a connected folder, pull from a connector, run code, write the Excel or PowerPoint output and check it. Chat answers a single prompt. Typical business use cases are recurring reports, document review sets, spreadsheet clean-up and research with a written deliverable.
Does Claude Cowork use our data for training?
On Team and Enterprise, Cowork conversation data is handled under the same commercial commitments as the rest of the plan's data and is not used to train Claude. Deleted tasks leave history immediately and backend storage within 30 days. Personal Pro and Max accounts follow the consumer policy instead, which is one reason company work belongs in a company workspace.
Can we see what employees do in Cowork?
Through the Compliance API on Enterprise and, per Anthropic's architecture overview, OpenTelemetry on Team and Enterprise, not through your endpoint protection, which cannot see inside the local VM or cloud sessions.
Discuss claude consulting for your business.
Tell us about your current systems, the result you need and your timeline. We will discuss the work, responsibilities and pricing before you decide on an engagement.