Tutorials
Claude Desktop: 5-Step Setup From MCP to Cowork
Set up the Claude desktop app from scratch — MCP extensions, Cowork agent, Computer Use, and power-user tips that'll save you hours.
March 25, 202612 min read
Anthropic's engineers built Cowork — the company's new AI agent for everyday users — in roughly a week and a half. And they built it using their own AI coding tool, Claude Code. That's either the most impressive demonstration of AI-assisted development or the most meta product launch of 2026.
Probably both.
The Claude desktop app has quietly become one of the most capable AI tools you can install on your machine. With MCP extensions, Cowork, and Computer Use now in the mix, it's a genuine productivity system — not just a chat window. This Claude desktop app tutorial walks you through the full setup: installation, MCP configuration, Cowork activation, and the features most people still haven't discovered.
What You'll Learn
This part's important — by the end of this guide, you'll have:
- A fully configured Claude desktop app on macOS or Windows
- MCP (Model Context Protocol) extensions connecting Claude to your files, databases, and tools
- Cowork running as your personal AI agent for non-coding tasks
- Computer Use enabled for hands-free screen interaction
- A handful of power-user tips that save real time
Prerequisites
Before we start, make sure you have:
- macOS 11+ or Windows 10+ — macOS gets features first, and Cowork is available on macOS and Windows as of March 25, 2026
- An Anthropic account at claude.ai
- A Claude Pro or Max subscription — the free tier works for basics, but the good stuff requires a paid plan
- A text editor for MCP config files (VS Code, Notepad++, even Text Edit)
- Node.js 18+ if you plan to use npm-based MCP servers
Step 1: Download and Install Claude Desktop
Head to claude.ai/download and grab the installer for your platform. The macOS version is a standard .dmg — drag it to Applications and you're done. Windows users get a conventional .exe installer.
First launch asks you to sign in with your Anthropic account. Do that, and you'll land in a clean chat interface that looks a lot like the web version. But don't be fooled — the desktop app can do things the browser never will.
The desktop app isn't just a wrapper around the website. It's a full client with file system access, MCP support, and agent capabilities the web version simply can't match.
So why bother with a desktop app when the web works fine? Two words: local access. The desktop client can read your files, run local MCP servers, and — with Cowork — operate directly on your documents. The web version is sandboxed. The desktop version isn't.
Step 2: Pick the Right Subscription Plan
Worth flagging: as of March 25, 2026, Anthropic offers three tiers:
| Plan | Price | Key Features |
|---|---|---|
| Free | $0/month | Basic Claude access, limited daily usage |
| Pro | $20/month | Priority access, higher limits, early features |
| Max | $100–$200/month | Everything in Pro plus Cowork, highest usage limits |
Cowork — the headline feature we'll cover in Step 4 — is available on Pro, Max, Team, and Enterprise plans. If you're here mainly for MCP and general desktop usage, Pro is plenty. But if you want the full agent experience, you'll need Max.
The model powering your conversations depends on your tier too. Max subscribers get full access to Claude Opus 4.6, which leads major benchmarks including MMLU and SWE-bench Verified, according to Anthropic's published evaluations. Pro users default to Claude Sonnet 4.6 — still highly capable across the same benchmarks.
Step 3: Configure MCP Extensions
This is where the Claude desktop app gets genuinely interesting. MCP — the Model Context Protocol — is Anthropic's open standard for connecting AI models to external tools and data sources. Think of it as USB ports for your AI: plug in a server, and Claude gains new abilities.
What MCP Actually Does
Without MCP, Claude can only work with what you paste into the chat window. With MCP, Claude can:
- Read and write files on your local machine
- Query databases directly
- Pull data from GitHub, Slack, or other APIs
- Run custom tools you define yourself
Setting Up Your First MCP Server
The configuration lives in a JSON file. On macOS, find it (or create it) at:
~/Library/Application Support/Claude/claude_desktop_config.json
On Windows:
%APPDATA%\Claude\claude_desktop_config.json
Here's a starter config that gives Claude read/write access to a folder on your machine:
JSON
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/yourname/Documents"
]
}
}
}
Replace /Users/yourname/Documents with whatever directory you want Claude to access. Restart the desktop app after saving — MCP config is only loaded at launch. You should see a small tools icon in the chat input area confirming the connection.
Popular MCP Servers Worth Adding
A few that are particularly useful:
- Filesystem — Read, write, and search local files
- GitHub — Browse repos, create issues, read pull requests without leaving Claude
- Postgres / SQLite — Query databases conversationally
- Brave Search — Give Claude real-time web search ability
You can find a growing directory of community MCP servers at modelcontextprotocol.io. And yes, you can run multiple servers simultaneously — just add more entries to the mcpServers object in your config.
MCP is the single feature that separates the Claude desktop app from every other AI chat client. Once you've configured it, going back to copy-pasting context feels like going back to dial-up.
Step 4: Set Up Cowork — Claude's Agent for Everyone
Don't skip this part. Here's the big one. According to VentureBeat, Anthropic released Cowork as a research preview in January 2026, and it represents a major shift in what the Claude desktop app can do for non-developers.
What Is Claude Cowork and How Does It Work?
Claude Cowork is an AI agent built into the Claude desktop app that lets non-technical users automate tasks like file organization, expense reports, and document research — much like how developers use Claude Code in the terminal. Available for Pro, Max, Team, and Enterprise subscribers on macOS and Windows, Cowork can open folders, read your documents, and complete multi-step tasks without you babysitting every click.
The origin story is pretty great. Anthropic noticed that developers were using Claude Code (their terminal-based coding tool) for decidedly non-coding work. According to Anthropic engineer Boris Cherny, users were doing vacation research, building slide decks, and cleaning up email — all through a coding tool. So Anthropic built a proper interface for that behavior.
Getting Started with Cowork
- Confirm you're on Max — Cowork appears for Pro, Max, Team, and Enterprise subscribers as of March 25, 2026
- Open Claude Desktop on macOS — look for the Cowork option in the sidebar or mode selector
- Point it at a folder — Cowork needs a working directory, similar to how Claude Code operates
- Give it a specific task — Instead of "organize my files," try something like: "Sort all PDFs in this folder by date, rename them with the format YYYY-MM-DD-description, and create a summary spreadsheet"
The key insight: Cowork works best when you give it a concrete goal and a clear scope. It's an agent, not a chatbot. It wants to do things, not just talk about them.
What Cowork Handles Well (and Where It Struggles)
Strong at:
- Processing batches of documents
- Creating structured outputs from messy inputs (think: receipts turned into expense reports)
- Research tasks that involve reading multiple files
- File organization, renaming, and categorization
Not great at yet:
- Tasks requiring real-time web access (unless you've got MCP configured for that)
- Anything needing pixel-perfect visual layout
- Multi-hour workflows — it's still a research preview, after all
Step 5: Enable Computer Use
Computer Use is Claude's ability to see your screen and interact with UI elements — clicking buttons, filling forms, and operating applications on your behalf. As of March 25, 2026, this remains in beta and requires explicit opt-in.
To enable it, check your Claude Desktop settings panel for a "Computer Use" or "Beta Features" toggle. When active, Claude can take screenshots of your desktop, interpret what's on screen, and execute mouse clicks and keyboard inputs.
A word of caution: Computer Use gives Claude significant control over your machine. (For context on how other companies approach this, see OpenAI's take on Computer Use.) Start with low-stakes tasks — let it fill out a simple form or reorganize some windows before trusting it with anything sensitive. And always supervise during your first few sessions.
Computer Use is like handing someone the keys to your car. Make sure they know how to drive before pointing them at the highway.
Tips and Common Pitfalls
After working through the full setup, here are the things that trip people up most:
1. MCP server not showing up? Restart the app. Every single time. The config file is only read at launch — this catches everyone at least once (and it'll probably catch you too).
2. Start with one MCP server. Don't add six on day one. Get filesystem working, confirm it connects, then expand. Debugging MCP problems with multiple servers running is no fun.
3. Use system prompts for recurring tasks. The desktop app lets you set custom instructions. "You're a financial analyst reviewing expense reports" is far more effective than re-explaining context every conversation.
4. Scope your Cowork tasks tightly. Don't say "organize my entire computer." Point it at one folder with one clear objective. Expand scope gradually as you build trust in the output.
5. Check which model you're running. The app defaults to the best model available for your tier. Opus 4.6 tops benchmarks like HumanEval and GSM8K — it handles complex reasoning tasks that Sonnet would struggle with. If you're on Max, make sure you're actually getting Opus.
6. MCP logs help when things break. On macOS, check ~/Library/Logs/Claude/ for diagnostic output. A missing comma in your JSON config will silently break everything — logs are where you'll spot it.
Verify Your Setup
Run through this quick checklist before calling it done:
- Claude desktop app launches and sign-in works
- Your subscription tier displays correctly in settings
- At least one MCP server connects (look for the tools indicator in the chat input)
- Ask Claude to read a file from your configured MCP directory — it should work
- Cowork appears in your interface (Pro, Max, Team, and Enterprise subscribers)
- Computer Use toggle is available under beta features
If anything fails, the usual fixes are: restart the app, double-check your JSON config for syntax errors, and confirm Node.js is installed if you're using npx-based MCP servers.
What's Next
Once you've got the fundamentals running:
- Build custom MCP servers — The protocol is open and well-documented if you have a specific workflow in mind
- Try Claude Code — If you're a developer, the terminal-based agent is where Anthropic's agent tech started. See our Goose vs Claude Code comparison for how it stacks up against alternatives
- Watch the MCP ecosystem — New community servers appear weekly for everything from email to calendar to CRM integrations
- Combine Cowork with MCP — The real power move is pairing Cowork's agent capabilities with MCP's external tool access. Give Claude the ability to search the web, read your files, and act autonomously. That's when things get exciting.
Sources
