Claude Code Desktop: Install, First Project, and the Real Learning Curve
šŸ’» code

Claude Code Desktop: Install, First Project, and the Real Learning Curve

Agentic Development with Claude

Part 1 of 1

Agentic Development with Claude

View all chapters →
Devesh Korde

Devesh Korde

November 17, 2026

šŸ“– 9 min read
#claude-code#claude#agentic-development#developer-tools
⚔ TL;DR
  • Claude Code Desktop is the Code tab inside the Claude desktop app. It includes Claude Code, so there's no Node.js or CLI to install, but it does need a paid plan.
  • A first session takes four choices: environment, project folder, permission mode, and model. Start with a small project you know well and review every diff.
  • Permission modes are the real learning curve. Manual, Accept edits, Plan, and Auto decide how much you hand over, and the right answer changes with the task.
  • The habits that pay off early: give context with @mentions and a CLAUDE.md, steer mid-run instead of restarting, comment on diffs, and watch your context window.

Week 2 of this series is about agents, and the first step is the least glamorous one: getting a working setup.

Claude Code Desktop is the Code tab inside the Claude desktop app. It gives you the same Claude Code engine as the terminal, with a graphical interface: a sidebar for sessions, visual diff review, an embedded terminal and file editor, and a browser pane for previewing your app.

Everything below comes from Anthropic's Claude Code docs, read on 2026-11-14. Desktop changes quickly, so where a screen doesn't match, check the version notes in the desktop reference.

What you need first

  • A paid plan. Claude Code requires a Pro, Max, Team, or Enterprise subscription. If clicking the Code tab prompts you to upgrade, that's why.
  • A project folder on your machine. Starting with a small project you know well is the fastest way to see what the tool does.
  • Git, if you want parallel sessions. Session isolation uses Git worktrees, which I cover below.

You don't need Node.js or the CLI. The desktop app includes Claude Code, so the Code tab works on its own.

Install

PlatformWhat to download
macOSUniversal build for Intel and Apple Silicon
Windowsx64 installer, with a separate ARM64 installer
LinuxA beta for Ubuntu and Debian, through apt or a .deb package

The steps are short:

  1. Install and sign in. Run the installer, launch Claude, and sign in with your Anthropic account.
  2. Open the Code tab. It sits at the top of the sidebar. If the tabs show only icons, hover to see the names.
  3. Fix the common first-click problems. A prompt to upgrade means you need a paid plan. A prompt to sign in online means finishing sign-in and restarting the app. A 403 error points to the authentication troubleshooting section in the docs.

One version note: the pane layout, integrated terminal, and file editor need Claude Desktop v1.2581.0 or later. If something is missing, use Check for Updates in the app menu.

Download, sign in, and open the Code tab
Download, sign in, and open the Code tab

Your first session

With the Code tab open, the row under the prompt box holds four controls. Set them before you send anything:

  1. Environment. Choose where Claude runs. For your first session, pick the top entry in the Local group, which runs on your machine. The other options are Cloud, SSH, and WSL on Windows.
  2. Project folder. Click the folder control, choose Open folder, and select your project.
  3. Permission mode. Decides whether Claude waits for approval. More on this below.
  4. Model. Pick one from the model picker. You can change it mid-session.

Then give it a small, concrete task. The quickstart suggests things like fixing a TODO comment, adding tests for a main function, or creating a CLAUDE.md with instructions for your codebase. That last one is a good first move, because it forces Claude to read the project and gives you something to correct.

The four controls under the prompt box
The four controls under the prompt box

Permission modes: the real decision

Setup takes minutes. This is the part that takes practice.

Permission modes control how much autonomy Claude has. You can switch at any time using the selector in that same row.

ModeWhat happens
ManualClaude asks before editing files or running commands. You see a diff and accept or reject each change.
Accept editsFile edits and common filesystem commands go through automatically. Other terminal commands still ask.
PlanClaude explores and proposes a plan without editing your source.
AutoClaude runs without routine prompts, while a background classifier checks actions like shell commands and network requests. Availability depends on your model and setup.
Bypass permissionsNo prompts, apart from a few actions no mode approves. The docs say to use it only in sandboxed containers or VMs.

Two details are easy to miss. A mode you pick is remembered per folder, except Plan, which applies only to the current session. And Desktop reads the same settings files as the CLI, so you can set a default with permissions.defaultMode.

My advice is to start in Manual, and use Plan before anything large. The docs make the same suggestion: map the approach first, approve it, then switch to Accept edits or Manual to execute.

This is where the co-driver idea earns its place. In Plan mode, you read the approach and decide. The architecture decision stays with you, and Claude does the typing afterward. If you hand over the decision as well as the keystrokes, you've swapped seats.

A dial from Manual to Auto, with the developer's hand on it
A dial from Manual to Auto, with the developer's hand on it

Habits worth building in the first week

Give context, not just instructions. Type @ and a filename to pull a file into the conversation, or attach screenshots, mockups, and PDFs from the + button. A CLAUDE.md in the project gives every session the same standing context.

Steer instead of restarting. You can interrupt with the stop button, or type a correction and press Enter to send it without stopping the running action. Claude reads it once the current action finishes.

Treat diff review as a skill. After Claude edits, click the line counts (for example +12 -1) to open the diff view. Hover a line and click the plus to leave a comment, then submit all comments at once. Claude reads them and revises. For a second opinion, type /code-review and Claude reviews the changes itself.

Use the panes. The integrated terminal (Ctrl+backtick) opens in your session's working directory and shares Claude's environment, so npm test sees the same files Claude is editing. The Browser pane previews your running app; Claude can start your dev server, take screenshots, and check its own changes. If your app uses a custom start command, edit .claude/launch.json.

Ask side questions. Press Cmd+; (Ctrl+; on Windows) or type /btw to open a side chat that sees the session's context without adding to it. Side chats aren't saved after you close the app.

A few shortcuts, with macOS keys (swap Cmd for Ctrl on Windows):

ShortcutAction
Cmd+NNew session
Cmd+Shift+DToggle diff pane
Cmd+Shift+BToggle Browser pane
Cmd+Shift+MOpen permission mode menu
Cmd+Shift+EOpen effort menu
Cmd+/Show all shortcuts

The effort menu connects to Day 4. Extended thinking is on by default in Desktop, which helps with complex tasks and uses more tokens, so effort is worth understanding before you hit usage limits.

Scaling up: parallel sessions

Once one session feels comfortable, open more. Click + New or press Cmd+N. Each session has its own context and changes, and you can view two side by side.

For Git projects, turn on Worktree before you start a session to run it in its own isolated copy of the project. It's off by default. Two practical notes from the docs:

  • Worktrees live in /.claude/worktrees/ by default.
  • Gitignored files like .env aren't copied into a new worktree unless you list them in a .worktreeinclude file.

Before archiving a session whose worktree has uncommitted changes, commit what you want to keep. Archiving can discard them.

If you open pull requests from Desktop, PR monitoring needs the GitHub CLI installed and authenticated. After that, Claude can watch CI and fix failing checks.

The learning curve

This is my reading of the docs, not a measurement. Five things look like they take real adjustment:

  1. Deciding how much to hand over. Permission modes put a dial in your hands. The right setting changes with the task, so expect to adjust it often at first.
  2. Managing context. When the context window fills, Claude summarizes and continues. You can trigger that earlier with /compact and inspect it with /context. Long sessions get less sharp as they accumulate, so smaller focused sessions tend to help.
  3. Understanding usage. Plan usage is shared across all your Claude Code surfaces, and the suggested next prompts that appear in the prompt box come from a background request that counts toward your limits. You can turn those suggestions off in Settings.
  4. Reading diffs as the main loop. The tool makes review easy, but it doesn't do it for you. The skill is reading each change as if a colleague wrote it.
  5. Environment snags. On Windows, Desktop doesn't read PowerShell profiles, so set PATH in your user or system environment variables and restart the app. On macOS and Linux, if an exported variable is missing in a command Claude runs, start a new session or restart the app.

Desktop or the terminal?

Desktop fits when you want visual diffs, panes side by side, and several sessions in one window. The terminal CLI fits scripting and keyboard-driven work. They aren't exclusive: Desktop reads the same settings files as the CLI, and you can install the CLI separately if you want claude in your terminal.

Next in the series, I look at the terminal side.

The short version

Install, sign in, open the Code tab, pick a small project, and start in Manual or Plan. Give Claude context, steer when it drifts, and read every diff.

Setup is the easy part. Knowing how much to hand over is the skill.


Sources

šŸŽ‰ You've finished Agentic Development with Claude

← Back to all articles

Related Articles