Everything you need to know to set up, automate, and master Claude's desktop agent.
Most AI tools are conversation tools. You type, they respond, and then you still have to go do the actual work yourself.
Claude Cowork works differently.
You describe what you want done, Claude figures out the steps, accesses your actual files, runs the work, and gives you finished outputs directly to your file system.
Think about your daily routine. You might have to pull data from 3 different places, reformat a report, send status updates to 5 clients. Cowork can run that whole block of work in a few minutes while you are doing something else.
1. Setup and installation
Getting started takes about 2 minutes.
- Step 1: Go to claude.ai/download and grab the Claude desktop app. It's available on both macOS and Windows with full feature parity on both platforms.
- Step 2: Open the app and log in. Look for the mode selector at the top of the interface. You'll see 3 options: Chat, Code, and Cowork. Click Cowork.
That's it. You're now in.
One thing to know: Cowork requires a paid Claude plan. It's available on Pro ($20/month), and Max ($100/month or $200/month for higher limits). The usage cost is higher than regular chat because multi-step agentic tasks are more compute-intensive.
A single Cowork session organizing files might use as much quota as dozens of regular chat messages.
Your first configuration step (don't skip this):
Go to Settings > Cowork > Edit next to Global Instructions. This is where you tell Claude who you are and how you work, and it applies to every session automatically.
EXAMPLE: Global Instructions
I'm a professor and AI content creator. My viewers are a mixed of general professionals and non-technical people who want to understand AI simply and practically. Written outputs should be clear, direct, and free of jargon. Reports in Word format unless I say otherwise. Default to landscape orientation for slides. My main projects are: social media content, newsletter posts, and YouTube videos on AI news and tutorials.
Now every task starts with Claude already knowing your context. You never re-explain yourself. You can also set folder-specific instructions, so your marketing folder gets different context than your research folder.
2. Your first task (and the Cowork loop)
Here's the core workflow that every Cowork task follows:
- You describe the outcome you want
- Claude plans the steps
- You review and approve
- Claude executes
- Finished work lands in your file system
Let me start with something almost everyone needs: organizing a messy folder (this actually was my first task when I used it for the very first time).
*Help me organize my Downloads folder. Scan the contents and
propose a plan: categories and folders to create, how files
should be sorted, any naming conventions to apply, and files
to flag for review. Show me the plan before making any changes.
Only proceed after I approve.*
Notice the last two sentences. Telling Cowork to show you the plan before executing is one of the most important habits you can build. Claude will come back with a structured proposal: categories like Documents, Media, Installers, Archives, the naming conventions it recommends, and a list of files it thinks you should look at. You review. You approve. Then it executes.
Here's a more advanced example that shows what Cowork can really do with local files:
*In this folder, I have meeting notes from client calls over the
last 3 months. Synthesize the key themes, recurring pain points,
and open action items into a single formatted Word document.
Organize it by client, then chronologically within each client.
Flag any action items that were mentioned more than once.*
That task would take a human assistant most of a day. Claude runs it in minutes, reading every document, pulling threads across them, and producing a clean Word report ready to share.
Key detail about file access: Cowork can only access the folders you explicitly grant it permission to. You control which folders Claude can read from and write to, either at the start of a task or through your Settings. This is direct local file access, not uploads or copy-paste. Claude reads your actual files and writes outputs back to your file system.
A WORD OF CAUTION: Claude Cowork can make mistakes, and because it has real access to your real files, those mistakes have real consequences. It can delete files you didn't want deleted, overwrite something you needed, or reorganize things in ways that break your workflow. Get in the habit of backing up important folders before running Cowork on them.
3. Connectors (linking Claude to your external tools)
Connectors are what turn Cowork from a local file assistant into something that touches your entire workflow. They link Claude to your external services via MCP (Model Context Protocol).
To browse them: Settings > Connectors > Browse Connectors.
You'll find connectors for Google Workspace (Calendar, Drive, Gmail), Slack, Notion, and a growing catalog of others.
There are 2 types: web connectors that run through browser-based APIs, and desktop extensions that run locally with deeper system access.
What makes connectors in Cowork different from connectors in regular Claude chat?
In chat, a Google Drive connector just shows you data.
In Cowork, that same connector can save data locally, combine it with your other files, run analysis on it, and push the output back to Drive, all in one task. The integration is active, not passive like qwith Chat.
*Pull the last 5 project status reports from my Google Drive.
Summarize the key milestones and blockers from each. Then draft
individual email updates for each project lead, using their name,
referencing the specific context of their project, and flagging
anything overdue. Save the drafts to Gmail, ready to review and send.*
That runs end to end: pulling from Drive, synthesizing across documents, writing personalized emails, and dropping them into Gmail. You review and hit send.
Claude in Chrome is another connector worth knowing about. When enabled, Claude can browse the web as part of a task without you switching windows.
It can pull data from websites, fill in forms, do competitive research, and combine what it finds online with your local files.
*Visit these 5 websites [list URLs], note any new blog posts,
product announcements, or pricing changes since last week,
and write a competitor intelligence summary saved to my
Drive folder in this format: [describe format].*
That runs as a real task touching the actual web and your actual file system.
4. Scheduled tasks (putting work on autopilot)
This is the feature that I love,it's like a digital employee that I've hired. You set a task, pick a schedule, and it just runs.
How to set one up: Type /schedule in any Cowork task. You can set tasks to run daily, weekly, monthly, or at a specific time. To manage them all, click "Scheduled" in the left sidebar where you can create, view, edit, pause, or delete any of them.
One limitation to know: scheduled tasks only run while your computer is awake and the Claude desktop app is open. If your machine is asleep when a task is due, Cowork will run it automatically once you wake your computer or reopen the app. This works best for tasks during your working hours, not overnight automation.
Here are 3 scheduled task prompts you can try:
SCHEDULED TASK: Monday morning priorities (weekly, 8am)
*Pull my calendar for the week from Google Calendar. Identify
my 3 busiest days. Draft a weekly priorities document that
includes my key meetings, deadlines, and recommended focus
blocks. Save it to my Desktop.*
SCHEDULED TASK: Friday file cleanup (weekly, 4pm)
*Scan my Downloads folder for anything added this week. Sort
new files into the right project folders based on filename
and content. Flag anything that doesn't clearly belong
anywhere and needs my attention.*
SCHEDULED TASK: Daily urgent email check (daily, 12pm)
*Check my Gmail for any messages marked urgent or flagged
today. Draft suggested responses for each one. Save as a
text file on my Desktop with the subject lines and my
draft replies.*
Something interesting that happens with scheduled tasks: after the first run, Claude rewrites its own instructions based on what it learned (which connectors it used, where it found data, what worked). The second run is usually better than the first because Claude has refined its own playbook.
5. Plugins (turning Cowork into a role-specific agent)
Plugins are bundles of skills, connectors, slash commands, and sub-agents packaged together for a specific role or workflow.
Instead of setting up each piece individually and re-explaining your process every session, you install a plugin once and Claude becomes a specialist from the first message.
How to install: Click the Customize menu in the left sidebar (this is where you see all your plugins, skills, and connectors in one place). Click Browse Plugins. Anthropic has open-sourced tons of plugins built by their own team, and new ones keep getting added.
Some examples of what's available:
The marketing plugin connects Claude to your content tools and gives you commands like /seo-audit, /content-brief, and /social-post. The sales plugin connects to your CRM with commands for prospect research and call prep. There are plugins for HR, legal, engineering, operations, financial analysis, and more.
Each was designed with practitioners in the relevant field, so the workflows and output formats match how that work actually gets done.
The slash command is super useful too. Once a plugin is installed, type / in the Cowork chat or click the + button to see all available commands. Select one, fill in the structured form, hit go, and Claude runs the full workflow.
You can also build your own custom plugins. If you have a workflow that's specific to your team or company, you can package it up so every team member gets the same agent experience from day one.
6. Power user tips and what's next
A few things to get real value out of Cowork:
Always ask for the plan first. End your prompts with "Show me the plan before making any changes. Only proceed after I approve." This takes 5 extra seconds and prevents mistakes.
Stack connectors in a single task. You can really take it a a whole level by chaining connectors, instead of using Google Drive or Gmail or Slack individually. Pull data from Drive, cross-reference with Slack threads, draft emails in Gmail, all in one prompt.
Write detailed global instructions. The more context Claude has about you, your role, your preferences, and your projects, the less time you spend correcting outputs. Spend 10 minutes on this once. It pays off on every task after.
Use folder-specific instructions for different projects. If you manage multiple clients or departments, each folder can carry its own context. Claude adapts automatically.
Build habits around scheduled tasks. Start with 1 or 2. A Monday morning planner. A Friday file cleanup. Once you see them running reliably, add more.
Cowork is still in research preview. Anthropic is shipping updates fast, including recent additions like Computer Use (Claude can control your keyboard and mouse when connectors aren't available), Dispatch (assign tasks from your phone), and a Projects feature for organizing work by local folders.
The trajectory here is clear: every knowledge worker gets their own custom agent.
The people who learn the patterns now will have a structural advantage over everyone who waits.
If you've made it this far, you'll enjoy my free newsletter: simplifyingAI.co. Every day, I send to your inbox a break down the most important AI news and practical AI tutorials like this.
