Claude Cowork is Anthropic's agentic workspace for knowledge work: a desktop app where Claude can actually read, edit, and create files in folders you choose, use your applications, and complete multi-step tasks from start to finish — rather than just telling you how. Anthropic describes it as bringing the power of Claude Code — its terminal coding agent — to everyone who works in documents, spreadsheets, and decks instead of code.
This guide explains how Cowork differs from a normal chat, how plugins turn it into a specialist, how a task actually runs, and who it is for.
How it's different from chat
In a regular Claude chat, the model can read what you paste and describe what to do, but it cannot touch your files or apps. Cowork removes that wall. Point it at a folder and it can open, edit, and save the files inside; let it interact with your screen and it can drive applications like Excel, PowerPoint, and Slack. The trade-off for that power is control: Cowork asks permission before it accesses each file or app, shows you what it is doing, and lets you stop it at any step.
What Cowork can do that chat can't
That combination — real file access, real app control, and multi-step execution — is what turns "Claude, explain how to build this model" into "Claude, build this model in my workbook and walk me through it." It is the difference between an assistant that advises and one that does the work.
Plugins: turning Claude into a specialist
The feature that makes Cowork more than a file-aware chatbot is plugins. A plugin bundles four things — skills, connectors, slash commands, and sub-agents — into a package that tells Claude how your role works: the best practices to follow, the tools and data to pull from, and the commands to expose. Instead of writing a detailed prompt for every task, you install a plugin and Claude already knows the domain.
Anthropic open-sourced eleven of its in-house plugins and built more for specific functions: human resources, investment banking and equity research, design, engineering, operations, private equity, wealth management, sales, marketing, and legal. Each ships role-specific slash commands — type /sales:call-prep or /data:write-query and the right workflow fires.
How a task runs
The pattern is goal-first. You state what you want; Cowork figures out the steps. Relevant agentic skills fire automatically, it coordinates sub-agents and tool calls for complex work, and it pulls from connectors — Google Workspace, Slack, DocuSign, FactSet, S&P Global, and more — when the task needs them, with native Excel and PowerPoint embedding for office work. Throughout, you stay in the loop: every app access is a permission prompt, and the whole run is visible and interruptible. (Underneath, those connectors largely speak MCP, the same open protocol the rest of Claude uses.)
Who it's for
Cowork is aimed at knowledge workers and, increasingly, large enterprises — the firms that have standardized on it read like a who's-who of professional services. It is available through Anthropic's higher-tier plans and is the surface to reach for when a task involves your real files and apps rather than a one-off question. If your work is mostly conversational, the standard Claude chat is enough; the moment you find yourself copying files in and pasting results back out by hand, that is the job Cowork was built to take over.
Cowork in practice
How organizations are actually putting Cowork to work:
- PwC bets the office of the CFO on Claude — pledging to certify 30,000 staff on Anthropic’s tools.
- Cowork moves into Big Law with 20+ integrations and a Microsoft 365 agent.
- A marketing plugin that runs campaigns through slash commands and an Ahrefs-to-Klaviyo stack.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.