Heads up: some links below are referral links. They cost you nothing extra and never change what we recommend. How this site makes money.
The short answer
Cowork runs the same agentic architecture as Claude Code, but inside the Claude desktop app with no terminal. You describe a task; it works through it across your files and connected apps and comes back with a finished deliverable.
Included with Pro, Max and Team. The mental shift that makes it work: brief it like a colleague, not like a search box.
How it differs from chatting with Claude
Ordinary Claude answers a question. Cowork does a job.
It reads and writes local files, works across connected apps like Slack and Google Drive, runs multi-step tasks to completion, and cites the actual files and messages it used. It can keep working in the cloud after you close your laptop, and you can start or redirect a task from your phone.
The unit of work is a deliverable — a report, a reorganised folder, a drafted response set — not a reply.
How to brief it
This is where most people go wrong. Cowork rewards the kind of instruction you'd give a competent new colleague, including the context you'd normally leave implicit.
Weak: "Sort out my invoices folder."
Better: "Go through the invoices folder, rename each file to YYYY-MM-DD-supplier-amount, move anything from before 2025 into an Archive subfolder, and give me a summary of the total per supplier for this year. Show me the renaming plan before you apply it."
Three things make the difference: a specific outcome, the rules to apply, and a checkpoint before anything irreversible.
What it's good at
- Multi-step file work — renaming, sorting, converting, extracting across many files at once.
- Pulling information together from several sources into one document, with citations back to where each piece came from.
- Recurring processes you can describe precisely — the weekly summary, the monthly reconciliation.
- Work that spans apps, where the answer lives partly in Drive and partly in Slack.
What it's not good at
- Vague creative direction. "Make this better" gives an agent nothing to aim at.
- Tasks needing judgement you haven't specified. It follows your rules; it can't guess the ones you didn't state.
- Anything you can't verify. If you couldn't check the result, you shouldn't delegate it.
Safety rules worth following
Ask for a plan before destructive operations. "Show me what you'll rename before you rename it" costs one extra step and prevents the bad afternoon.
Work on copies the first few times, until you trust it with a given task.
Treat file contents as data, not instructions. A document from outside your organisation can contain text aimed at manipulating an agent. Be cautious pointing Cowork at untrusted material.
A good first task
Pick something real, bounded and reversible. Good candidates:
- "Read these five documents and give me a one-page summary of where they disagree."
- "Go through this folder and tell me which files are duplicates — don't delete anything yet."
- "Turn this messy CSV into a clean spreadsheet with headers and consistent date formats."
Each has a clear finish line and nothing irreversible, which is exactly what you want while working out how much to trust it.
Common questions
Is Claude Cowork free?
Do I need to know how to code?
Can Cowork access my files?
Does it keep working if I close my laptop?
What's the difference between Cowork and Claude Code?
Cowork is included with Pro
No separate subscription — if you have a paid Claude plan, it's already there.