Claude Cowork

What it is

Claude Cowork is Anthropic’s agentic work surface for people who don’t write code — a tab in the Claude desktop app next to Chat and Code where you hand Claude a goal rather than a prompt, and it works across folders you’ve granted, connected SaaS apps, a built-in browser panel, and, on Pro and Max, your actual screen. It reached general availability on 9 April 2026, after a macOS public beta in January and Windows in February. It now runs on macOS, Windows (x64 and arm64), Linux and ChromeOS, with the web version rolling out and mobile in beta since 7 July 2026.

The nearest entries in this catalog are Lindy and Dust, agent platforms sold to operators. The difference is that Cowork is not a builder: no canvas, no node graph, no agent to configure before you get an answer. You describe the outcome and steer mid-run.

Why it shows up in ops stacks

The reason is distribution before capability. Cowork is included on every paid Claude plan at no extra charge, so an ops leader at a company that already bought Claude seats for engineering now has an agent surface for the finance analyst and the recruiting coordinator — no new procurement cycle, no new vendor security review, no new line item. Every other tool in this category requires a separate purchase first.

Three capabilities do the actual work:

• Scheduled tasks run in the cloud. Typing /schedule in a task, or using the Scheduled entry in the sidebar, sets a recurring job that executes on Anthropic’s servers whether or not your laptop is open. This is the piece that turns Cowork from a faster assistant into unattended throughput.
• Connectors rather than screen-scraping. Microsoft 365, Google Drive, Slack and Amplitude are official, and the connector browser in Settings adds Notion, Gmail and others. Connector-mediated access is faster and more reliable than the computer-use path, and unlike computer use it is available on every plan.
• Dispatch. One continuous thread you can reach from your phone. Claude routes the request itself — engineering work into Claude Code, knowledge work into Cowork — and sends a push notification when it finishes or needs an approval.

Pricing reality

Cowork has no price of its own. What you buy is a Claude seat, and every paid tier includes it:

• Pro — $20/mo, or $17/mo billed annually
• Max — $100/mo at 5× Pro usage, $200/mo at 20×
• Team — $25/seat/mo, $20 on annual, two-seat minimum; premium seats at $125/$100 for 5× the usage
• Enterprise — $20/seat plus usage, with organization-level feature and model controls

The tier list is not the real number. Anthropic publishes no token quotas, only multipliers relative to Pro, and usage is metered twice: a session cap that resets every five hours and a weekly cap on a fixed reset assigned to your account. That budget is pooled across Claude Code, claude.ai chat and Cowork — a Cowork scheduled task and an engineer’s coding session draw from the same allowance. Teams running Cowork daily against real workloads land on Max or on premium Team seats, not on Pro. Budget the $20 seat for evaluation and the $100–$200 band per heavy user for production.

Best for

Ops leaders in RevOps, Legal Ops, Recruiting and Customer Success at companies already paying for Claude, who need to hand a non-engineer a recurring document-and-spreadsheet job that runs unattended: a weekly pipeline-hygiene report built from CRM exports in a Drive folder, a monthly diff across a folder of vendor contracts, a Monday candidate-pipeline digest assembled from ATS exports and Slack threads. The scoped case where it wins outright is recurring, file-heavy work with a stable output format, owned by one named person, over data that already sits in folders or connectors that person is entitled to see.

Not for

Cowork acts as you — your seat, your folder grants, your connector permissions. There is no per-agent service identity, no agent credential to rotate, and no audit subject distinct from the human who owns the session. If the requirement is a registry of agents with their own lifecycle and their own audit trail, this is the wrong layer and buying more seats will not fix it; that is the Microsoft Agent 365 and Salesforce Agentforce shape of problem. It is also the wrong tool for high-volume structured pipelines. If the job is “process 5,000 rows the same way every night,” an n8n flow or a direct API call against Claude is cheaper and far more predictable than an agent reasoning its way through each row.

Versus the alternatives

• Microsoft 365 Copilot — the incumbent by seat count inside ops organizations; since 22 April 2026 its agentic mode is generally available in Word, Excel and PowerPoint. Pick it when the work lives in Office documents on SharePoint, the governance conversation is about tenant-level policy, and you want the agent inside the application your team already has open rather than in a separate desktop client.
• ChatGPT Work — launched 9 July 2026 on a GPT-5.6 agent, with a connector directory of more than 1,400 apps, on web, mobile, and a macOS desktop app that brings Chat, Work and Codex together. Pick it when connector breadth is the binding constraint, or when the organization has already standardized on OpenAI. Its plan-then-execute-with-approvals model is the closest here to Cowork’s.
• Manus — the fastest-growing entrant in the segment, at $100M ARR eight months after launch and a $125M revenue run-rate by December 2025. Pick it for open-ended research-and-build tasks where you want a finished artifact from a one-line brief. The reason it is not a default: ownership is unresolved — Meta’s $2B acquisition was blocked by Chinese regulators in late April 2026, and a buyback by the company’s earlier backers was reported in June 2026. Do not put a governed, recurring workflow behind a company whose cap table is in dispute.

Watch-outs

• Computer use is Pro and Max only — Team and Enterprise plans do not have it, and where it exists it is macOS and Windows only; the Linux build gets files and connectors but no screen control. This catches buyers who evaluate on a personal Pro seat and then roll out on Team seats, at which point the demo stops working. Guard: before buying seats, split the target workflows into the ones that need screen control and the ones a connector can reach. Re-scope everything you can onto connectors; for the remainder, keep those specific users on individual Max seats and budget them separately.
• The session sandbox limits network reach, not your data. Anthropic’s own safety documentation is direct about this: a session runs in an isolated, temporary cloud environment that cannot reach your home or company network and is destroyed when the session ends, but that isolation “doesn’t limit what Claude reads or does” through the folders and connectors you granted — and computer use has no sandbox between Claude and what is on your screen. Guard: grant access at the narrowest directory that holds the job’s inputs, never a home directory or the root of a synced drive; keep credentials, payroll and financial records out of any granted folder; and on Enterprise, switch on skill scanning and disable Cowork web search for the groups that do not need it.
• Scheduled tasks keep running after the person who wrote them stops watching. They execute in the cloud whether or not the machine is on, and Anthropic names web content as the primary prompt-injection vector — so a task that reads a public page is running untrusted text on a schedule with your permissions. “Automatically approve” still screens each action for safety before it runs, but it does not stop an injected instruction mid-task. Guard: keep scheduled tasks on manual approval for anything that writes, and never combine a web-reading step and a write-capable connector in the same scheduled run. Review outputs on a set cadence and pause any task nobody has read in a month.