benchr
Most AI agents you've heard about are aimed at developers. Cowork is the one pointed at everyone else. The pitch is a single sentence: give it a goal, and Claude works across your computer, your local files, and your applications to hand back a finished deliverable. Not a chat reply you then act on. The actual thing, sorted folder, drafted report, filled spreadsheet, done.
What makes that possible is where it lives. Cowork runs on the desktop, in the Claude Desktop app, because that's where knowledge work happens: in local files, folders, and the apps you use all day. You grant Claude permission to read, edit, and create files in folders you choose, and it moves between those sources, synthesizes across them, and completes the task without you steering each step.
How a Cowork task runs
The shape of every job is the same, and it's worth seeing laid out, because it's what separates an agent from a chatbot.
Plain language, outcome-shaped. "Turn these receipts into an expense spreadsheet." "Draft the quarterly update from my notes."
It reads the local files and folders you've allowed, pulls from connected apps, and carries out the multi-step work without you coordinating each move.
It returns the actual output, and by design the consequential decisions stay with you. It does the work; it doesn't quietly send the email.
That last step is the one to underline. Cowork is built so it completes tasks but leaves the calls that matter to you. It's an agent with a hand on the wheel, not an autopilot, which is the right posture for software that's touching your real files.
The rollout, in order
Cowork didn't arrive all at once, and where it is today is worth pinning down precisely.
| Date | Milestone | Who got it |
|---|---|---|
| Jan 12, 2026 | Research preview | Claude Desktop, macOS, Max plan |
| Jan 16, 2026 | Wider preview | Added Pro plans on macOS |
| Apr 9, 2026 | Generally available | macOS and Windows, all paid plans |
So as of today it's no longer a macOS-only research preview. It's generally available on both macOS and Windows for Pro, Max, Team, and Enterprise subscribers. If you have a paid Claude plan, it's already in your desktop app.
What it connects to
Local files are the start, not the limit. Cowork ships with connectors for the tools knowledge work runs on: Google Workspace, covering Drive, Gmail, and Calendar, plus Docusign, FactSet, Slack, and Chrome for web research, among others. On Team and Enterprise plans, admins get real controls: managing who can use it, capping spend, and building private plugin marketplaces that bundle skills, connectors, and sub-agents into one installable package. That's the difference between a personal helper and something an organization can deploy.
Where it fits
The cleanest way to understand Cowork is by what it isn't. It runs on the same agentic architecture as Claude Code, but Claude Code is a terminal tool for developers, and Cowork is the desktop version for everyone else: researchers, analysts, operations, legal, finance. The model muscle underneath comes from Anthropic's frontier line, the same family covered in the Claude Opus 4.8 review, and the security-focused extreme of that agentic capability is the restricted model in the Claude Mythos explainer.
If you've been reading about agents and wondering when they'd do something useful for non-engineers, this is the clearest answer yet. The broader state of that shift is in AI agents, eighteen months in, and since a Cowork task can chew through a lot of tokens crossing your files and apps, it's worth pairing with cutting your token bill if you lean on it daily.