The Complete Guide to Claude Cowork: Your AI Desktop Agent
Everything developers need to know about Anthropic's agentic desktop tool — from first setup to advanced automation.
Claude Cowork is Anthropic’s answer to a question developers have been asking since Claude Code took off: what if non-technical teammates could use this too?
Launched in January 2026 as a research preview and promoted to general availability in April 2026, Cowork takes the same agentic architecture behind Claude Code — file access, multi-step execution, tool integration — and wraps it in the Claude Desktop app with a UI anyone can use. No terminal required.
If you’ve been watching the agentic AI space heat up with Microsoft’s Copilot Cowork, Google’s Gemini Agent Mode, and OpenAI’s Operator, Cowork is Anthropic’s entry: a desktop-native agent that reads, edits, and creates files on your machine while you step away.
This guide covers everything: what Cowork is, how to set it up, what it can actually do, and the advanced features that make it worth paying attention to.
What Is Cowork, Really?
In regular Claude chat, you type a message and get a response. Claude can’t touch your files. It can’t produce a finished deliverable. You copy-paste back and forth.
Cowork changes that relationship. You point Claude at a folder on your computer, describe what you want done, and Claude works through it — reading files, writing documents, creating spreadsheets, organizing folders — while you watch or walk away.
Under the hood, Cowork runs inside a sandboxed virtual machine on your desktop (using Apple’s Virtualization Framework on macOS). Your files are mounted into this container, which means Claude gets real filesystem access while staying isolated from the rest of your system.
Think of it this way:
Chat = you ask, Claude answers
Cowork = you assign, Claude delivers
Who Is It For?
Cowork was built for knowledge workers who aren’t developers but have the same kinds of repetitive, file-heavy tasks that Claude Code handles so well for engineers.
Anthropic reports that the majority of Cowork usage comes from outside engineering — operations, marketing, finance, legal, and research teams. The pattern they’re seeing: people aren’t handing Claude their core judgment work, but rather the assembly work that surrounds it. Project updates, collaboration decks, research sprints, data organization.
That said, developers benefit too. If you’re already a Claude Code user, Cowork gives you a friendlier interface for non-coding tasks: organizing documentation, compiling reports from scattered notes, processing data files.
Requirements and Setup
What You Need
Claude Desktop app (macOS or Windows) — download at claude.com/download
A paid Claude plan: Pro ($20/month), Max ($100 or $200/month), Team, or Enterprise
Active internet connection throughout your session
Desktop must stay awake and app must stay open while Claude works
Windows Users
If you’re on Windows, you can download a readiness checker from the Claude help center to verify your system supports Cowork before installing.
Getting Started
Download and install (or update) the Claude Desktop app
Open the app and find the mode selector
Click the Cowork tab to switch from Chat mode
You’ll land in the Tasks view — click + New task to begin
Your First Cowork Task
Starting a task is straightforward:
Click + New task
Optionally attach a folder by clicking the folder icon — this gives Claude access to read and write files in that directory
Type your prompt describing what you want done
Click Let’s go
Claude will analyze your request, propose a plan, and start executing. You’ll see progress indicators showing what it’s doing at each step.
A Good First Task
A classic starter is file organization:
Look at my Downloads folder. Scan the contents and propose a plan:
- Categories/folders to create
- How files should be sorted
- Any naming conventions to apply
- Files to flag for review or deletion
Show me the plan before making changes.Notice the last line — asking Claude to show the plan before acting. This is important. Cowork gives Claude real write access to your files, so building in checkpoints is good practice.
How Execution Works
When Claude receives your task, it:
Analyzes the request and creates a plan
Breaks complex work into subtasks when needed
Runs code and shell commands in the isolated VM
Coordinates multiple workstreams in parallel if appropriate
Loops you in before anything significant
You can steer at any point — redirect, refine, or change approach mid-task.
What Cowork Can Do: Core Capabilities
File Operations
Cowork’s foundation is filesystem access. Claude can read, create, rename, move, and delete files in folders you grant access to. This covers:
Sorting and renaming chaotic folders
Deduplicating files
Extracting text from PDFs and images
Converting between file formats
Document Creation
Cowork has built-in skills for producing office files: Word documents (.docx), PowerPoint presentations (.pptx), spreadsheets (.xlsx), and PDFs. These aren’t chat-style outputs — they’re real, editable files saved to your filesystem.
Example prompt:
This folder contains our Q3 operating expenses spreadsheet.
Analyze the data and create a Word document report covering:
- Total annual spending
- Breakdown by category with percentages
- The three largest expense categories
- Month-over-month trends
Format it professionally with clear sections, tables,
and an executive summary.Research and Synthesis
Claude can combine local files with web research to produce comprehensive outputs. Point it at a folder of notes, articles, or data files and ask it to synthesize a report, identify themes, or build a briefing document.
Data Processing
Feed Claude receipt screenshots, expense reports, or raw data files, and it can extract, categorize, and structure the information into clean spreadsheets.
The Customization Stack: Skills, Connectors, and Plugins
This is where Cowork gets powerful. The customization system has three layers that build on each other.
Skills
Skills are instruction files that tell Claude how to do specific kinds of work. They govern output format, tone, methodology, and quality standards.
Cowork ships with built-in skills for common file types (pdf, docx, pptx, xlsx). But custom skills are where things get interesting.
You can create a skill for anything: your brand voice, your report format, your review framework. Once installed, the skill applies automatically whenever Claude does relevant work.
To create a custom skill:
Use the built-in Skill Creator — it interviews you about what you want, then generates a properly structured skill file you can install and toggle on or off.
To browse and install skills:
Go to Customize in the left sidebar
Click the Skills tab, then the + button
Select Browse skills to see available options
For more on browsing all customization options, see Browse skills, connectors, and plugins.
Connectors
Connectors link Claude to external services and data sources. In regular Claude chat, connectors pull information. In Cowork, they gain filesystem access — meaning a connector that fetches data from an external service can now save that data locally, or use local files as input.
The connector catalog includes hundreds of options: Gmail, Google Calendar, Slack, Notion, Asana, Linear, Jira, AWS, Salesforce, and many more. Both web connectors and desktop extensions are available.
To browse connectors:
Go to Settings > Connectors > Browse connectors. You’ll see two tabs: Web and Desktop extensions.
Plugins
Plugins are the highest-level building block. A plugin bundles skills, connectors, slash commands, and sub-agents into a single installable package tailored to a specific job function.
Anthropic open-sourced 11 starter plugins on GitHub covering roles like sales, finance, legal, marketing, HR, engineering, design, and operations. Each comes pre-configured with relevant skills and connectors. There are also specialized financial services plugins for investment banking, equity research, private equity, and wealth management.
To install a plugin:
Switch to the Cowork tab
Click Customize in the left sidebar
Click Browse plugins
Click Install on your chosen plugin
After installing, Claude walks you through customizing the plugin — adjusting skills and connectors to match your specific workflow and tools. For more details, see Use plugins in Cowork.
Building your own plugin:
You can build a plugin from scratch using Plugin Create, a built-in tool. Describe what you want — even a sentence is enough — and Claude assembles the skills, connectors, and commands into a plugin package.
Alternatively, start from any Anthropic template and modify it. Plugins are just markdown files and JSON, so no code or infrastructure is needed.
Projects: Persistent Workspaces
Introduced in March 2026, Projects solved Cowork’s biggest early limitation: every session started from zero. You had to re-explain context and re-upload files each time.
A Cowork Project is a dedicated workspace with:
Instructions — tone, formatting, rules that apply to every task in the project
Context — a local folder, linked chat project, or URLs for Claude to reference
Memory — Claude remembers context from previous tasks within the project
Scheduled tasks — recurring automation scoped to the project
Conversation history — a persistent thread that doesn’t bleed into other work
Think of Projects as separate desks for separate jobs, instead of one cluttered desk for everything.
Creating a Project
Open Cowork
Go to Projects in the left sidebar
Click + and choose your method:
Create from scratch
Import from an existing Claude chat project
Use an existing folder
Name it, set instructions, attach files, and click Create
Memory Within Projects
Memory is scoped per project. What Claude learns in your marketing project doesn’t leak into your finance project. Over time, Claude picks up patterns — it might notice recurring themes in weekly reports or remember your preferences for how data should be formatted.
Scheduled Tasks: Automation on Autopilot
This is the feature that changes Cowork from a tool you use to a tool that works for you.
You write a prompt once, pick a cadence (daily, weekly, monthly), and Claude runs it automatically. No code, no APIs, no cron jobs. For full details, see Schedule recurring tasks in Cowork.
Setting Up a Scheduled Task
Open Cowork and start a new task (or open an existing one)
Type
/schedulein the chat inputClaude walks you through defining the task and cadence
Once confirmed, the task runs on schedule
Scheduled tasks have access to everything regular tasks do: connectors, plugins, skills, sub-agents.
Practical Examples
Here are some high-value scheduled tasks:
Morning briefing:
Check my email and calendar for today.
Summarize any messages that need a response
and list my meetings with relevant context.
Save the briefing as a markdown file in my
/daily-briefs folder.
Weekly report compilation:
Every Friday, compile data from the /project-updates
folder into a formatted status report.
Include progress against milestones,
blockers flagged this week, and key metrics.
Recurring file cleanup:
Every Sunday, organize my Downloads folder by type
and delete anything older than 30 days.
Important Limitations
Tasks only run while your computer is awake and the Claude Desktop app is open
If your machine is off during a scheduled run, the task gets skipped (but shows up in task history and auto-runs when you reopen the app)
Scheduled tasks consume more usage than regular chat
Dispatch: Remote Control from Your Phone
Dispatch lets you assign tasks to your desktop Cowork agent from your phone. Your desktop stays active and does the work; your phone is just the remote.
Setup
Open Cowork on your desktop
Click Dispatch in the left side panel
Follow the setup flow to connect your mobile app
Enable file access and keep-awake settings
What You Can Do
From your phone, you can:
Ask Claude to pull data from a local spreadsheet and compile a summary
Have Claude search your Slack and email, then draft a briefing document
Request a formatted presentation built from files in your Drive
Tell Claude to organize files in a specific folder
Your desktop must be active for this to work. Claude uses the same connectors, plugins, and file access you’ve already configured.
Safety Note
Dispatch creates a chain where mobile instructions trigger real actions on your desktop — including reading, moving, or deleting files. A manipulated instruction or unexpected command could cascade. Only use this if you’re comfortable with what your agent could do, not just what you intend it to do. For more, see Use Cowork safely.
Sub-Agents: Parallel Execution
When Claude gets a complex task, it can spin up sub-agents — parallel workers that each handle a piece simultaneously.
Instead of reading 20 files one at a time, Claude might create five sub-agents that each process four files at once. One agent researches a company while another pulls CRM data and a third scans Slack for context.
Sub-agents run in parallel or in sequence, each with its own focus and fresh context window. You don’t configure this directly — Claude decides when parallel execution makes sense based on the task complexity.
Enterprise Features (GA Launch, April 2026)
The April 2026 general availability launch added features specifically for enterprise deployment:
Role-based access controls — admins manage who can access Cowork and what they can do
Group spend limits — control costs across teams
Private plugin marketplaces — admins create curated plugin collections for their org
Auto-installed plugins — push approved plugins to all team members
Usage analytics — admin dashboard plus Analytics API for deeper analysis
OpenTelemetry observability — monitoring and tracing for enterprise deployments
Granular MCP permissions — fine-grained control over which connectors and integrations are available
Tips for Getting the Most Out of Cowork
1. Start with a CLAUDE.md File
Create a markdown file in your working folder with context about yourself, your role, your preferences, and your project. Claude reads this automatically and applies it to every task.
# About Me
- Role: Product Manager at [Company]
- I prefer concise, data-driven reports
- Use metric system for all measurements
- Default output format: Google Docs-compatible .docx
# Current Projects
- Q2 launch planning
- Customer feedback analysis
2. Be Specific About Outcomes, Not Steps
Bad: “Look at these files and do something useful”
Good: “Analyze the CSV files in this folder. Create a summary spreadsheet with one row per client, columns for total revenue, number of transactions, and average deal size. Sort by revenue descending.”
3. Build in Checkpoints
Since Cowork has real file access, ask Claude to show you the plan before executing. Use phrases like “show me the plan before making changes” or “propose your approach first.”
4. Use Folder Instructions
You can set folder-specific instructions that kick in whenever you’re working in that folder. These persist across sessions and complement your global preferences.
5. Batch Related Work
One Cowork session uses significantly more compute than a regular chat (roughly 20x by some estimates). Batch related tasks together rather than running many small sessions.
6. Test Before You Automate
Run a task manually first. Once you’re happy with the output quality, then convert it to a scheduled task with /schedule.
Cowork vs. the Competition
Here’s how Cowork stacks up against the other desktop/agent tools as of April 2026:
Claude Cowork is the only tool with full local file access (via a sandboxed VM), multi-step agentic execution, a plugin ecosystem, scheduled recurring tasks, mobile dispatch, and enterprise-grade controls (RBAC, spend limits). It runs on both macOS and Windows.
ChatGPT Desktop offers limited file access and basic multi-step execution, but lacks scheduled tasks, plugins, mobile dispatch, and has only limited enterprise controls. Available on macOS and Windows.
OpenAI Operator supports multi-step execution but only for web-based tasks — it can’t touch your local files. No scheduled tasks, no plugins, no mobile dispatch, and limited enterprise controls. Web only.
Google Mariner handles multi-step execution but is confined to Chrome. No local file access, no scheduled tasks, no plugins, no mobile dispatch, and no enterprise controls.
The short version: nobody else does what Cowork does for local file management. Operator excels at web automation but can’t touch your files. ChatGPT Desktop is still primarily a chat window. Mariner is Chrome-only.
Known Limitations
Be aware of these before you go all-in:
No team sharing for Projects — Cowork Projects are local to your machine. For shared workflows, you still need browser-based Claude Projects.
No cloud sync — everything happens locally. You can’t start a task on your laptop and continue on your desktop.
Computer must stay awake — if your machine sleeps or the app closes, active tasks stop and scheduled tasks get skipped.
Higher usage consumption — Cowork sessions use significantly more tokens than regular chat. Pro plan users will hit limits faster.
Computer Use is macOS only — the feature where Claude controls your mouse, keyboard, and desktop apps only works on Mac currently.
Memory is project-scoped only — Claude doesn’t retain memory across standalone Cowork sessions (only within Projects).
Pricing Quick Reference
Cowork is included on all paid Claude plans:
Pro ($20/month) — Good for getting started, but expect to hit usage limits with heavy Cowork use since each session consumes significantly more tokens than regular chat.
Max 5x ($100/month) — 5x the usage limits of Pro. This is the recommended tier for regular Cowork users who run tasks daily.
Max 20x ($200/month) — 20x the usage limits. Built for power users running heavy automation and multiple scheduled tasks.
Team (per-seat pricing) — Everything in Max plus admin controls for managing access across your team.
Enterprise (custom pricing) — The full governance layer: role-based access controls, analytics dashboard, Analytics API, OpenTelemetry observability, and private plugin marketplaces.
Getting Started Today
Here’s a 15-minute path to your first productive Cowork session:
Install or update the Claude Desktop app from claude.com/download (2 min)
Switch to the Cowork tab and start a new task (1 min)
Attach a folder — start with your Downloads folder or a project folder (1 min)
Run a simple task — file organization, document summarization, or data extraction (5 min)
Install a plugin that matches your role from the built-in marketplace (3 min)
Set up one scheduled task — a morning briefing is a great first choice (3 min)
From there, build up: create a Project for your most common workflow, add connectors to your key tools, and create custom skills for your specific standards and preferences.
The speed at which Cowork has evolved — from research preview in January to enterprise GA in April — signals that Anthropic is treating this as a core product, not an experiment. If you’re building on Claude or advising teams that use it, Cowork is worth understanding now. For a deeper dive, Anthropic offers an official Cowork training course covering the task loop, plugins, and steering techniques.
If you found this guide useful, share it with a colleague who’s drowning in file management.