Claude Cowork: The Complete Guide (2026)

What is Claude Cowork?

In traditional Chat mode, you maintain a conversation: you ask, Claude responds, and you orchestrate each step. Cowork fundamentally changes this paradigm. Instead of directing each micro-step, you describe a final outcome and let Claude plan, execute, and deliver the result autonomously.

Cowork runs locally on your machine through the Claude Desktop application. Operations execute inside an isolated virtualized Linux environment using Apple's native virtualization framework (AVF) on macOS. This architecture ensures Claude cannot access directories you haven't explicitly authorized -- the boundaries are architectural, not software-based.

Setup and First Configuration

Prerequisites

• A Mac running macOS (Windows support announced but not yet available as of March 2026)
• Claude Desktop from claude.com/download
• A paid subscription: Pro ($20/mo), Max ($100-200/mo), Team, or Enterprise

Installation Steps

1. Download and install Claude Desktop from the official site
2. Log in with your Claude account
3. Switch to Cowork mode using the tab at the top of the app (Chat | Code | Cowork)
4. Select a working folder -- create a dedicated directory like ~/Cowork-Projects
5. Write your first task description (describe an outcome, not a question)

Global Instructions

Global instructions persist across all sessions and apply to every task. Access them in Settings > Cowork > Edit next to "Global instructions." Include your role, preferred formats, tone, data handling rules, and any recurring organizational standards.

Writing Effective Cowork Prompts

A good Cowork prompt answers three fundamental questions:

1. What does "done" look like? (the deliverable)
2. What context does Claude need? (files, background info)
3. What constraints can't it guess? (format, tone, length, audience)

Weak prompt: "Make me a sales report."

Effective prompt: "Using sales-Q1-2026.csv in my working folder, create a 3-page PDF report. Include: an executive summary (5 lines), a table of sales by region, a monthly trend chart, and 3 actionable recommendations. Professional tone, intended for the board of directors."

Skills: Structured Reusable Workflows

Skills are structured, reusable workflows that Claude invokes automatically when relevant or manually via the / command. Each Skill represents a specific capability: summarizing documents, cleaning datasets, extracting structured data, or generating reports.

Creating Custom Skills

• Skill Creator wizard: an interactive tool that generates the Skill file for you
• Manual creation: write a Markdown file in .claude/skills/ in your working folder
• Community import: download and customize Skill templates from GitHub

Multiple Skills can work together. A task involving research followed by document creation will draw on both a research Skill and a writing Skill seamlessly.

Plugins: Pre-Configured Solution Packages

Plugins bundle multiple Skills, connectors, and sub-agents into pre-configured solution packages. For example, a "Sales Outreach" plugin might include Gmail connectivity, Google Sheets integration, Tavily web search, and embedded Skills for prospect analysis and email drafting.

Access plugins via Customize > Browse plugins in the sidebar. Anthropic provides official plugins for sales, finance, legal, marketing, HR, engineering, design, operations, and data analysis. After installing, click Customize on any plugin to adapt its Skills and connectors to your specific workflow.

Connectors: Integrating Your Tools

Connectors let Cowork interact with external services via the Model Context Protocol (MCP). Available connectors include Google (Gmail, Drive, Calendar), Microsoft (Outlook, OneDrive, M365), Slack, Notion, Figma, GitHub, Linear, DocuSign, and Tavily MCP for web search.

Connect a service via Customize > Connectors, authenticate with OAuth, and the connector becomes available for your tasks. For services without pre-built connectors, deploy a custom MCP server and connect Cowork to it.

Claude follows a tool selection hierarchy: direct connectors first (fastest), then browser automation, then Computer Use as a last resort.

Scheduled Tasks: Recurring Automation

Scheduled tasks transform Cowork into a continuous automation system. Create them by typing /schedule in an existing task or through the Scheduled tab in the sidebar. Specify a name, description, prompt, and frequency (hourly, daily, weekly, or weekdays).

Common use cases: daily email briefings, weekly status reports, recurring competitive intelligence, periodic file cleanup, and regular data processing. Tasks leverage all of Cowork's capabilities including connectors, plugins, Skills, and Computer Use.

Important: scheduled tasks only run while Claude Desktop is open and your Mac is awake. Missed tasks execute automatically when the app reopens.

Computer Use: Direct Desktop Control

Computer Use lets Claude interact with your actual desktop -- clicking, typing, navigating applications -- when connectors and browser automation aren't enough. Enable it in Settings > General > Computer Use.

Key safety measures:

• Claude asks for explicit permission before using Computer Use
• Add sensitive apps (banking, healthcare, crypto) to the blocklist
• Deletion protection requires your approval before removing files
• Computer Use operates outside the VM, so stronger safeguards apply

Best for straightforward, predictable interactions with applications that lack APIs. Not ideal for complex multi-step workflows.

Dispatch: Remote Control from Your Phone

Dispatch, launched March 2026, turns your phone into a remote control for Cowork. Send a task from your mobile, Claude executes it on your desktop, and results appear on your phone.

Setup

1. Enable Dispatch in Claude Desktop
2. Open the Claude mobile app
3. Scan the QR code to pair devices

Each Dispatch task spawns an independent sub-agent on the desktop. You can submit multiple tasks in parallel. Your Mac must stay awake for execution.

Projects: Persistent Workspaces with Memory

Projects, introduced in March 2026, solve Cowork's biggest limitation: no memory between sessions. A Project bundles files, instructions, scoped memory, scheduled tasks, and dedicated Skills into a self-contained workspace.

As you work within a Project, Claude automatically generates and stores memories: your preferences, decisions made, patterns identified, and key project details. Memory is stored locally, isolated per project, and persists across interruptions. You can view, edit, and delete memories at any time.

Tips and Best Practices

Optimize Your Prompts

• Describe outcomes, not steps: say "Deliver a 5-page PDF report" not "Open an editor, write a title..."
• Provide rich context: include example deliverables, style guides, reference docs
• Specify invisible constraints: output format, tone, target audience, max length
• Use folder instructions: create .claude/instructions.md for recurring project folders

Manage Token Consumption

Cowork tasks consume significantly more tokens than Chat conversations. Monitor usage in Settings, prefer Chat for simple tasks, break complex tasks into sub-tasks, and invest in global instructions to avoid repeating context.

Security Best Practices

• Use a dedicated working folder -- never grant access to your entire home directory
• Add sensitive applications to the blocklist
• Prioritize connectors over screen automation
• Watch for unusual patterns and stop tasks if something looks wrong

Current Limitations

• macOS only (Windows coming soon)
• Machine must stay awake for task execution
• Variable reliability on complex multi-step workflows
• High token consumption for autonomous tasks
• Local-only memory -- no cross-device sync
• Occasional connector instability with some services

Getting Started Checklist

1. Start simple: a document synthesis or file organization task
2. Configure your global instructions from day one
3. Connect your main tools (Gmail, Google Drive, Slack)
4. Create your first Project for persistent memory
5. Explore plugins for your field
6. Add a scheduled task for a recurring workflow

The key to success with Cowork is thinking in terms of outcome delegation, not step direction. Describe what you want, provide the necessary context, and let Claude orchestrate the execution.