Claude Cowork for Business: What It Does and Where It Breaks

Most business use of AI still happens in a chat window. Someone types a question, reads the answer, copies it somewhere useful, and repeats that a dozen times a day. The tool never sees the company's files, never remembers what happened last week, and cannot act on anything by itself. Claude Cowork is Anthropic's attempt at the next step. You hand over a whole task, and Claude works through it on your own files and systems while you get on with something else.

That shift raises a different set of questions. What can it reach, what does it cost to run, what happens when the laptop is shut, how much can an IT department control, and where does it fall down. This article answers those from Anthropic's own documentation. Launch coverage has aged badly in places, and much of it still repeats limitations that no longer exist.

The gaps get as much space here as the features, since they usually decide whether Cowork suits a business. The largest one affects any company buying more than a couple of seats, and the marketing does not mention it.

What this article covers

Claude Cowork for business, in short
How Cowork runs on your machine
Claude Cowork for business pricing
Connecting Cowork to your tools
Local files
Scheduled tasks
Dispatch and computer use
The team problem nobody mentions
Admin and IT controls
Where Cowork breaks
Sources

Claude Cowork for business, in short

Anthropic describes Cowork as bringing Claude Code's capabilities to knowledge work beyond coding. It reached general availability across all paid plans in April 2026 and comes with every tier from Pro upward at no extra subscription. The desktop app carries the full version, with web and mobile in beta.

A single task can span your local folders, your calendar, your email, a CRM and a browser, and Claude reports each step it took. You approve file changes as they happen, and deletions always ask.

Three boundaries sit around all of it. Claude reaches only the folders and tools you nominate, and with permissions enabled it asks before acting. An administrator on Team or Enterprise can also switch Cowork off for the whole organisation or for particular groups.

How Cowork runs on your machine

A local session runs the agent loop natively on the device. Code execution sits inside an isolated virtual machine using the platform's own hypervisor, which means Apple Virtualization.framework on macOS and Hyper-V on Windows. A remote session runs the agent loop and the code execution in a temporary sandbox on Anthropic's servers. That sandbox dies when the session ends.

Anthropic states the remote sandbox has no access to your network by default and cannot reach private, internal, link-local or cloud-metadata addresses. Outbound traffic passes through a proxy the sandbox cannot reconfigure or bypass. Only allow-listed destinations are reachable, and the tokens it holds expire within hours. A remote session needing a local file routes the request back through the desktop app, reaching only the folders you connected.

Supported platforms are macOS, Windows, and Linux in beta on Ubuntu and Debian. On Linux, Cowork runs tasks in a virtual machine. That needs hardware virtualisation through KVM, 25GB of free disk for the workspace image, and at least 8GB of RAM. Computer use is unavailable there.

Claude Cowork for business pricing

Team now starts at two seats where it once needed five, bringing it within reach of very small operations. Enterprise has a self-serve tier that removes the mandatory sales call. Current plan pricing runs as follows.

Free. $0, single seat. Chat, connectors and file creation. Cowork is not included
Pro. $17/month annually or $20 monthly. Adds Cowork, Claude Code, unlimited projects and Microsoft 365
Max 5x. From $100/month. Five times the Pro usage allowance and higher output limits
Max 20x. Twenty times the Pro allowance, for continuous heavy workloads
Team standard. $20/user/month annually or $25 monthly, two to 150 seats. Adds SSO, central billing and admin controls, with no model training by default
Team premium. $100/user/month annually or $125 monthly. Five times the standard allowance
Enterprise self-serve. $20 per seat plus usage at API rates. Adds spend controls, role-based access, SCIM, audit logs and HIPAA-ready handling
Enterprise sales-assisted. Custom pricing, negotiated by seat count

Cowork consumes considerably more of your allowance than chat does. It draws from the same pool as Claude chat and Claude Code, so heavy use in one reduces what the others have. Anthropic puts this plainly, saying complex multi-step tasks are compute-intensive and require more tokens. Price a rollout on the tier above whichever one looks adequate.

Connecting Cowork to your tools

Connections run on the Model Context Protocol, an open standard Anthropic published and then handed to the wider industry. MCP gives Claude and external software a shared language. Any service supporting it connects once, after which Claude reads from it, writes to it and acts on it inside a task.

The official directory lists tested connectors sorted by use case and by capability. You can see whether one is read-only, read and write, or interactive before installing it. Independent developers have published MCP servers for tools Anthropic has not built, and these install as plugins with a click. Quality there varies with the developer. For anything touching company data, the vendor's own connector or a directory entry is safer.

A custom connector needs only a web address and an access key. That puts an internal database, a bespoke CRM or a proprietary platform within reach, given some developer time.

Plugins, skills and connectors now share a single Customize menu, which puts everything a user can reach in one place.

Local files

Cowork reads, writes and edits files on your own file system, not a cloud copy. Files stay where they are and you point Claude at the folder. That covers reviewing a folder of contracts for unusual payment terms, or consolidating a directory of CSV exports into one summary. It also renames files to a convention and finds every document mentioning a particular client across several directories. The same mechanism builds a full presentation from local files without leaving the conversation.

Supported formats cover Word, PDF, Excel, PowerPoint, CSV, JSON, Markdown, code files and Jupyter notebooks. File operations wait for your approval, and deleting anything always asks.

Scheduled tasks

Scheduled tasks run hourly, daily, weekly, on weekdays, or manually. They execute remotely, so they keep their cadence with the computer asleep and the desktop app closed. In March that was the ceiling on Cowork, and always-on automation meant using the API.

The exception is a task that touches local files or local applications, which runs on your machine and therefore needs it awake.

A morning briefing makes a straightforward first build. One task checks Gmail at 6am on weekdays, pulls the day's calendar, and reads a Notion board for overdue items. It drafts the briefing and leaves a follow-up email written but unsent for approval. Pipeline monitoring works the same way. A daily task queries the CRM for deals with no recent activity. It posts the list, with values and last contact dates, into a Slack channel.

Dispatch and computer use

Computer use has moved out of API-only beta into Cowork itself. That changes the picture for any business running software never built to integrate. Claude operates on-screen applications the way a person does, opening them, clicking, typing, filling spreadsheets and navigating dashboards. Older ERP platforms, government portals and Citrix environments carry no connector and no webhook. The screen is the only route into them. A staff member who spends part of every week pulling the same figures out of a legacy interface has a candidate for this.

Anthropic still labels computer use a research preview. In practice it handles structured, repeatable workflows with predictable interfaces, and degrades on anything less stable.

Dispatch handles the phone side. You message Claude from mobile and the work runs on your desktop against the local files and connectors there. It notifies you on completion, or when it needs a decision. Both apps need to be current, and the desktop machine awake with the app open. The work happens on your computer, and the phone only assigns and collects. Scheduled tasks need no machine at all, so the two features have different requirements. Dispatch runs as one continuous thread and remains in beta on Pro and Max.

Any tool that reads outside content and then acts on your machine faces prompt injection. Instructions hidden inside a web page or a document can try to make Claude do something nobody asked for. The approval prompts are the defence. Disabling them to save a few clicks is a poor trade in any business setting.

The team problem nobody mentions

Anthropic's documentation states that for members of Team and Enterprise plans, Cowork projects do not support project sharing. Projects are desktop-only and stored locally, with no cloud sync for project data. Memory is scoped to each project, so what Claude learns in one does not carry to another. What it remembers about you in chat does not carry into Cowork at all.

Each person's Cowork therefore sits on their own machine, holding projects, memory and accumulated context that stay on that machine. No single database records what the team has been doing, and no colleague can pick up where another left off. Ten people using Cowork in one company are running ten separate installations.

The practical cost falls on whoever maintains standards. A house style, an approvals process or a client naming convention goes into instructions inside every project belonging to every user. Each copy then has to be kept in step by hand. Departing staff take their Cowork context with the laptop, and two people working the same account each see only their own. Sharing on Team and Enterprise is limited to live artifacts. That covers an individual output, without the session or project behind it.

Shared organisational memory has to be built separately. One route is a connector into a system everyone can reach. Another is a shared document store that each project's instructions point at. Cowork will not provide it and does not claim to.

Admin and IT controls

Single sign-on with domain capture removes the separate Claude account, and SCIM provisioning adds and removes users through Okta or Microsoft Entra. Role-based access controls scope which capabilities each group reaches, covering Cowork itself, Claude Code, web search and connectors. Connector permissions work per tool, granting read access while write access stays blocked.

Admins can also curate a private plugin marketplace, pre-approving plugins, auto-installing them for particular groups and hiding them from others. Private GitHub repositories work as a source, currently in private beta. OpenTelemetry streaming sends Cowork events to a SIEM. Enterprise adds spend controls, data retention controls, IP allowlisting and network access controls.

Two gaps in that picture matter for regulated work. Local session history sits on the user's own computer. Anthropic states plainly that this data falls outside their standard retention policies, and that admins cannot centrally manage or export it. Compliance API coverage is less clear, because the documentation contradicts itself. The Team and Enterprise article says Cowork via mobile and web is captured. The architecture overview says Cowork activity is not captured in audit logs, the Compliance API, or data exports. A compliance position that turns on the answer needs it in writing from Anthropic rather than from either page.

OpenTelemetry streaming is the control to configure first, since it is the one visibility mechanism the documentation does not dispute. Without it, agentic activity stays outside the monitoring a company already pays for.

Where Cowork breaks

Some of what follows is a hard stop. The rest is work Cowork will attempt and do poorly. That does more damage, since a refusal is visible and a plausible wrong answer is not.

No shared memory across a team. The largest gap on any multi-seat deployment, set out in full two sections above.

It consumes your allowance fast. One agentic task spawns sub-agents, tool calls and file operations, and burns through far more than an equivalent conversation. The pool is shared with chat and Claude Code.

Judgement work still needs a person. Retrieval, collation, reformatting and following a defined procedure all go well. Deciding which of two defensible answers suits a particular client, contract or regulator does not. A person should read financial figures, regulated information and client-facing output before it goes anywhere, because Cowork formats a wrong answer as neatly as a right one.

Computer use is the weakest component. Still a research preview, it degrades on dynamic layouts and unusual controls. Treat it as the last resort for systems with no API, and keep someone checking the output rather than letting an unmonitored process depend on it.

Compliance visibility is incomplete, and Anthropic's own pages disagree on how much of it the Compliance API captures.

Three further limits have not come up above.

Native connectors for older HR and ERP platforms remain thin in the directory, so those systems need a community server or a custom connector
Nothing leaves Cowork without a configured integration, so it cannot email a client a report unless a mail connector exists and is permitted
Plugins carrying local MCP servers, and live artifacts, work on the desktop app only

Cowork earns its cost where staff lose hours moving information between applications and rebuilding the same reports by hand. Getting value out of it depends on specifying that work clearly enough to hand it over.

Sources

All sources verified 25 July 2026.

Get started with Claude Cowork. Anthropic's own description, requirements, and named limitations
Claude Cowork architecture overview. Local and remote execution models, sandbox network boundaries, audit-log coverage
Use Claude Cowork on Team and Enterprise plans. Local data storage, admin visibility limits, Compliance API wording
Organize your tasks with projects in Claude Cowork. Project sharing, local-only storage, project-scoped memory
Use Claude Cowork on web, desktop, and mobile. What each surface can and cannot do
Schedule recurring tasks in Cowork. Cadences, and remote execution with the machine asleep
Assign tasks to Claude from anywhere in Cowork. Dispatch requirements and limits
Claude Cowork product page. Stated capabilities and the permission model
Claude pricing. Plan tiers, prices and seat limits
Model Context Protocol. The open standard behind connectors
Claude connectors directory. Available connectors by use case and capability
Install Claude Desktop. Platform support, including Linux beta scope