Getting Started with Claude Code: Your First Session

The Claude Code CLI is the most direct way to start. Install it via npm, point it at a repository, and start typing instructions.

From your working directory (a Git repository):

1npm install -g @anthropic-ai/claude-code2cclude init3cclude start

The first cclude init command creates a .claude/ directory with a default configuration. The cclude start command launches an interactive session. From there, you can type instructions like Fix the race condition in payment-processor.js or Add type safety to this API response schema, and Claude Code will:

1. Read and search your actual repository 2. Propose edits to files 3. Ask for permission if the edit touches a sensitive path or tool 4. Execute shell commands to run tests or build steps 5. Report what changed and what still needs work

The session persists in memory within that terminal, so context from earlier steps is available to later steps — if you ask for a bug fix and then say run the test suite to verify, Claude Code remembers both the fix and the original problem statement.

If you spend most of your time in VS Code, the Claude Code extension brings the agent loop directly into your editor. Install from the VS Code marketplace, sign in with your Anthropic account, and open any workspace.

Once installed, you'll see a Claude Code sidebar on the left. Open it, paste in your instruction, and Claude Code will read files from the workspace, make edits inline (you'll see diffs before confirming), and run terminal commands in the integrated terminal.

The key difference from CLI: edits are shown as side-by-side diffs before applying, and you can reject or accept individual changes rather than all-or-nothing. This is the interface most teams use for their daily workflow because it keeps you in the same editor where you're already working.

The Claude Code Desktop app is a standalone executable (no web browser needed) that works on macOS, Linux, and Windows. Download from the official Anthropic website, install, and open your project folder from File → Open.

The Desktop experience is visually similar to VS Code — a left sidebar for instructions, a main area showing results and diffs, and a lower panel for terminal output. The main difference is that Desktop is self-contained: it does not depend on VS Code being installed, and it uses your system's default shell (zsh, bash, PowerShell, etc.) rather than embedding a terminal.

The Desktop app is often the first choice for new users who have not yet set up a CLI development environment, and it is especially useful on Windows where terminal tooling varies.

Open https://claude.dev in any browser, connect your cloud storage or Git repository (GitHub, GitLab, Bitbucket), and start a session. The web version is useful for quick edits when you don't have the Desktop or VS Code extension installed, and it's the easiest way to onboard non-engineering team members (product managers, doc writers) to small, scoped tasks.

The trade-off: the web version has a shorter history window and slightly higher latency than Desktop or CLI, because the filesystem interaction happens over a network boundary. For deep refactoring across fifty files, Desktop or CLI is faster. For a quick hotfix or docs update, the web interface is convenient.

Claude Code on mobile (available via the Claude app on iOS and Android) is designed for supervision and quick edits when you're away from a desk. It works identically to the web version but is optimized for touch and small screens.

Typical mobile use case: you get a notification that a build failed; you open Claude Code on your phone, skim the error log, and ask Claude Code to apply a quick fix. The app confirms edits visually (because small-screen diffs are hard to read, the app shows a summary before applying), runs the fix, and shows you the result.

Mobile is not a replacement for Desktop or CLI for serious development work — the screen size makes reviewing large diffs difficult. It is, however, a critical tool for on-call engineers and remote teams who need to respond to production issues without access to a laptop.

The best first tasks for Claude Code are narrowly scoped, have verifiable success criteria, and teach you how to work with the agent loop without overwhelming it. Here are three starting points:

Task 1: Fix a class of bugs. If you have an open bug ticket that you know is caused by a specific pattern (e.g., "missing null check on this object in three places"), that's a perfect first task. Give Claude Code the ticket, point it to the file, and ask it to find and fix every instance. Claude Code will search the repository, show you each fix, and run your test suite to verify.

Claude Code asks for permission before it does three things: (1) editing files outside a designated safe zone, (2) executing commands that might have side effects (especially database commands or deployments), and (3) accessing secrets from environment variables or .env files.

When you initialize a repository with cclude init or open it in Desktop, the .claude/settings.json file sets defaults. By default, Claude Code is conservative: it allows reads everywhere but requires explicit permission for writes, and it never touches secrets without asking.

You can loosen these rules in .claude/settings.json if you trust Claude Code on your team. For example:

1{2  "permissions": {3    "readOnlyPaths": [],4    "safeWritePaths": ["src/**", "tests/**"],5    "allowSecretAccess": false6  }7}

With this config, Claude Code will automatically write to anything under src/ or tests/ without asking, but will still ask before touching a secret. These rules are per-repository, not global — you can be strict in production code and permissive in internal tools.

After your first task, you've learned the basic loop: describe what you want → Claude Code reads files → shows diffs → applies changes → runs tests → reports results. From here:

For deeper understanding: Read How Claude Code Works to learn about context compaction, subagents, and why the agent loop is designed the way it is.

For extending Claude Code: Add MCP (Model Context Protocol) servers to connect to databases, APIs, or monitoring tools. Start with MCP Integration: Connect Claude Code to External Tools.

For advanced workflows: Explore Claude Code Workflow to learn about skills, hooks, and how to integrate Claude Code into your team's CI/CD pipeline.

For troubleshooting: Check Claude Code Permissions if you hit unexpected permission denials, and Claude Code Context Window if your sessions start running out of memory.