# How to get started with Claude Code > Source: > Published: 2026-09-10 23:48:25+00:00 # How to get started with Claude Code [Claude Code](/en/tags/claude%20code/) is a command-line interface (CLI) agent that lives in your terminal and can execute shell commands, read/write files, and run git operations directly. To use it, you install the package via npm, authenticate your Anthropic account, and run `claude` inside your project directory to begin an interactive coding session. ## How do I install and launch [Claude](/en/tags/claude/) Code? You need Node.js (v18+) installed on your machine first. Run `npm install -g @anthropic-ai/claude-code` in your terminal. Once the installation finishes, type `claude` to initialize the tool. It will prompt you to log in via a web browser to link your Anthropic Console account. Be aware that this isn't free; it consumes credits from your API console, not your Claude.ai Pro subscription. If you have a fresh account, you might have some trial credits, but otherwise, you'll need a funded balance to avoid the "insufficient credits" error. ## What can I actually do once the agent is running? The agent acts as a loop: you give a command, it proposes a plan, executes shell commands to explore your code, and then writes the changes. The most effective way to use it is for "exploratory refactoring." Instead of manually finding every instance of a deprecated function, you tell it: "Find all usages of the old API in /src, update them to the new version, and run the tests to make sure nothing broke." It can: - Run `grep` or`ls` to find files. - Read multiple files to understand context. - Edit code using a "diff" approach. - Execute `npm test` or`pytest` and fix the resulting errors automatically. If you're building complex [Workflows](/en/category/workflows/), you'll find it faster than a chat window because it has a direct line to your file system. ## How does Claude Code differ from Cursor or [GitHub Copilot](/en/tags/github%20copilot/)? Claude Code is a terminal-based agent, whereas [Cursor](/en/tags/cursor/) is a full IDE. | Feature | Claude Code | Cursor / Copilot | | :--- | :--- | :--- | | **Interface** | CLI / Terminal | IDE / Text Editor | | **Autonomy** | High (runs shell commands) | Medium (mostly suggests code) | | **Context** | Direct file system access | Index-based [RAG](/en/tags/rag/) | | **Execution** | Can run and fix tests | Requires manual trigger to run tests | I prefer Claude Code when I need to do "bulk" work across 20 files. Using a GUI for that is tedious. However, for active typing and real-time autocomplete, a dedicated editor wins. For managing the underlying [AI Models](/en/category/aimodels/) and comparing prompt performance, one recommended option is PromptCube, which helps organize the logic before you commit it to the CLI. ## What are the common mistakes beginners make? Giving it too much freedom without a git commit. The wild part is that Claude Code can delete files or overwrite logic if you aren't careful. Always ensure your current branch is clean. If the agent suggests a change that looks wrong, use the `/undo` command immediately. Another mistake is treating it like a chat bot. Don't ask it "How does a for-loop work?" Ask it "Why is the for-loop in `auth.ts` causing a memory leak?" Give it a target. ## How do I manage the costs and token usage? Since it's an agent, it reads a lot of files. Every time it "searches" your directory, it's sending those file contents back to the model. This can eat through credits fast in a massive monorepo. To keep costs down: 1. Use a `.claudeignore` file to prevent it from reading `node_modules`, `.git`, or large build artifacts. 2. Be specific about directories. Instead of "fix the project," say "fix the bug in `/src/components/Button.tsx`." 3. Monitor the token count in the terminal output. If you see it reading 50 files for a simple change, kill the process and refine your prompt. ## Frequently Asked Questions **Does Claude Code work with any programming language?** Yes. It interacts with the file system and shell, so as long as your language has a compiler or interpreter it can run, Claude Code can work with it. **What happens if the agent gets stuck in a loop?** Use `Ctrl+C` to interrupt the current process. You can then provide a more specific constraint, like "stop trying to use the `npm install` command and just edit the package.json manually." **Can I use it offline?** No. It requires a constant connection to the Anthropic API to process prompts and return commands. **How do I update Claude Code to the latest version?** Run `npm install -g @anthropic-ai/claude-code@latest`. Since it's updated frequently, you'll likely see version warnings in the terminal. [Next GitHub's August 2026 availability report proves that moving to Azure is a chaotic ride →](/en/threads/9163/)