What is Claude Cowork?
Claude Cowork is where you hand Claude real work. Instead of answering your messages one at a time, it takes on a complex, multi-step task, plans how to get there, and executes it across the files and tools you choose - returning a finished document, deck, spreadsheet, or piece of research for you to review. It is the same Claude you already talk to, wearing a knowledge worker's tools instead of a chat window.
The point is a shift in what you ask for. In a normal chat, you describe how to do something and Claude describes it back. In Cowork, you define the goal and Claude figures out how to reach it - opening files, running the steps, and doing the formatting and fetching you would otherwise do by hand. You say what you need, not how.
Cowork is where you hand Claude real work. You define the goal. Claude figures out how to get there.- anthropic.com/product/claude-cowork
Cowork is built on the same agentic architecture that powers Claude Code, Anthropic's coding agent. Claude Code brings that engine to software engineering in the terminal; Cowork brings it to knowledge work beyond code, with no terminal required.
Chat answers, Cowork acts
The clearest way to understand Cowork is against the chat you already know. In regular chat, Claude responds to your messages but cannot reach your files directly - it works with whatever you paste in or upload, and it stops at describing the work. In Cowork, Claude has permission to read, edit, and create files in the folders you specify, so it can actually complete the task rather than explain how to do it.
The two are not separate apps. Chat and Cowork share one home: you start both from the same message box. Find the composer, select Cowork to hand off a task, or Chat to go back to a normal conversation. The switch is the whole difference in surface - the model underneath is the same.
How a task runs
When you give Cowork a goal, it does not answer in one shot. It runs a loop you can watch and steer:
Plans the work. Claude analyses your request and lays out a plan, breaking a complex job into subtasks when it needs to.
Works in parallel. It coordinates multiple workstreams at once where that helps - drafting while it also researches and organises, rather than doing each in sequence.
Runs in an isolated environment. In beta, sessions run remotely on Anthropic's servers, separate from your computer and network. Code and shell commands Claude writes run inside that sandbox, and when a task needs something local - a file on your disk, your browser - it reaches it through the Claude Desktop app on that computer.
Delivers for review. Finished outputs land back in your session, where you can preview and download them. You review polished work rather than raw steps.
Throughout, Claude shows each step - the files it opens, the tools it uses, the choices it makes - so you keep visibility into what it is doing and can redirect it when it matters, or let it run on its own.
What it produces
Cowork is measured by the artifact it hands back, not the conversation. The typical outputs are the deliverables of an office day:
Slide decks
Turn scattered exports and notes into a formatted, review-ready presentation - the slides built, not described.
Formatted documents
Briefs, memos, and reports written and laid out in the format you asked for, ready to sign off.
Spreadsheets
Data wrangled, cleaned, and organised into a spreadsheet instead of a wall of pasted numbers.
Synthesised research
Sources gathered and distilled into a single organised summary, with the legwork already done.
The throughline mirrors Claude Code's: the unit Cowork operates on is a task, not a question. "Give me a weekly Monday deck with recommended actions from these campaign exports" is a typical brief - and it ends with a deck, not with instructions for making one.
Files, tools, and connectors
Cowork can do real work because it acts on real inputs, under permissions you set. Three things make that possible:
- Folders you choose. Claude can read and write files only in the folders you connect - it cannot reach anything else. Network access follows the egress settings you configure, and deleting anything needs your approval.
- Connectors. Cowork plugs into the tools your work already lives in. Anthropic has shipped connectors for services like Google Drive, Gmail, DocuSign, and FactSet, so a task can pull from your drive, your inbox, or your data source directly.
- Plugins and skills. Reusable plugins and agent skills encode institutional knowledge and workflows - financial analysis, engineering, HR - so Cowork follows your team's conventions rather than generic defaults. Enterprise admins can even run a private plugin marketplace.
The result is an agent that knows about your specific environment - your files, your inbox, your systems of record - not just generic public knowledge.
Steering, schedules, and projects
Handing off a task does not mean losing control of it. Cowork is built to be steered and to keep running when you step away:
- Approval gates. With permissions enabled, Claude shows its plan and waits for your approval before anything significant. It takes over your screen only when it has to, and you can follow along - or redirect it - from anywhere.
- Scheduled tasks. Set a task to run on a cadence - daily, weekly, or monthly - and it runs unattended. Attach your campaign exports, schedule a weekly job, and come back to a fresh deck every Monday.
- Projects. Organise related tasks into persistent, self-contained workspaces with their own files, links, instructions, and memory, so recurring work carries its context forward.
Because sessions run remotely in beta, work continues if you close your laptop. You can start a task at your desk, check in from your phone, and open the same session from any surface.
Where it runs, and who it is for
Cowork runs on Claude Desktop for macOS and Windows, and is rolling out on web and mobile in beta - the same Claude wherever you work. It is available on the paid plans: Pro, Max, Team, and Enterprise. For organisations, admins can manage feature access, control spend, and track Cowork usage across the org.
Who is it for? Anyone whose day is knowledge work rather than code. Anthropic frames the ambition plainly: engineers came to treat Claude Code as a tool they could not work without, and the company expects knowledge workers to feel the same way about Cowork. It goes beyond the basic chatbot interface by bringing real-time context from your connected tools, plus the admin controls enterprises expect.
When to reach for it
The three surfaces of Claude divide cleanly by the shape of the job:
Ask a question
- Quick answers, drafting, thinking out loud
- No file access needed
- Best for: research, writing, planning
Hand off a task
- Multi-step work across your files and tools
- Returns a finished deck, doc, sheet, or brief
- Best for: knowledge work you want done, not described
Reach for Cowork when the work is a real, multi-step task that touches your files and ends in a deliverable - and when you would rather review finished work than walk each step yourself. Reach for plain chat when you just need an answer, and for Claude Code when the task is software engineering in a repository. For a direct, axis-by-axis breakdown of that last call, read Claude Code vs Claude Cowork.