Claude Code: Where to Start — A Beginner's Guide

In short

Claude Code is Anthropic's terminal AI agent: it reads your project's code, edits files and runs commands right in the console. To get started, install the CLI (curl -fsSL https://claude.ai/install.sh | bash on macOS/Linux, or npm install -g @anthropic-ai/claude-code), run claude inside your project folder, sign in with a Claude Pro/Max subscription or an API key, run /init — and you can give it your first task in plain text. After that you connect MCP servers (claude mcp add) and skills so the agent can see current documentation, your database and a browser. Below: step by step, with a command table.

What is Claude Code, and what do you need to use it?

Claude Code is a development agent that lives in the terminal rather than in a separate editor. You describe the task in words and it finds the right files itself, makes the edits, runs the tests and shows you the result. What sets it apart from a chatbot is that it actually acts on your file system instead of just handing you snippets to copy.

What you'll need to start:

• A terminal (on Windows, PowerShell or Windows Terminal will do).
• Paid access: a Claude Pro, Max, Team or Enterprise subscription or an Anthropic API key from the console. The free Claude.ai tier does not include Claude Code — an honest caveat that trips up nearly every beginner.
• For the npm install, Node.js 22+ (the native installer doesn't need Node.js).

If you're still choosing between Claude Code, Cursor and VS Code, see the breakdown in Claude Code vs Cursor vs VS Code — it covers how transferable the skills are and when each one is more convenient.

How do you install Claude Code?

At the time of writing (July 2026) the recommended method is the native installer: it drops in a platform binary and doesn't depend on Node.js. The npm package works too and installs the same binary under the hood.

Important: install the npm package without sudo — installing via sudo breaks permissions and gets in the way of updates later. After installing, check the version:

claude --version

If the command returns a version number, the CLI is in place. It updates itself, but you can force it with claude update.

How do you sign in to Claude Code?

Go to any project folder and run:

claude

On first launch the CLI will prompt you to sign in. There are two paths:

1. Subscription (Pro/Max). A browser opens and you confirm the login through your Anthropic account via OAuth. The simplest option if you already have a subscription.
2. API key. The ANTHROPIC_API_KEY environment variable is used for direct API access with a key from the Anthropic console. You pay per token. You can set or switch the key later through /config.

The /login command lets you change the sign-in method at any point. That's it — authentication is a one-time step, and the CLI remembers your choice from then on.

Your first project and the basic commands

Go into your project folder (for example, cd my-app) and run claude. The first thing to do is run /init — it creates the CLAUDE.md file.

CLAUDE.md is the project's memory file: the agent reads it at the start of every session and keeps the rules, the stack and the code conventions there. The tidier that file, the less often the agent does something you didn't expect. After that you just write the task in human language, for example: "add a /health route that returns status 200 and the app version."

The quality of the answer depends heavily on how the task is framed. If the agent constantly "doesn't get it", start by reading how to brief an agent so it actually understands — that saves more time than any setting.

Commands in Claude Code start with a slash. Here's the minimum worth memorizing right away:

Two commands save you most often: /clear, when the conversation has gone sideways and the agent is tangled in old context, and /compact, when the session is long and the context is about to overflow. A detailed breakdown of the CLI's capabilities and practical techniques is in the Claude Code card in the arsenal.

How do you connect MCP servers?

MCP (Model Context Protocol) is the standard by which the agent gets access to external tools: documentation, databases, a browser, GitHub. Without MCP the agent only knows your files; with MCP it can, for example, read a library's current docs or run real queries against your database.

Connecting takes one command. The basic syntax:

claude mcp add <name> <transport and parameters>

There are three transports: stdio (a local process or npm package), http (cloud tools, recommended) and sse (deprecated — take http if you have the choice). Servers have a visibility scope:

• the default is local: the server is visible only to you and only in the current project;
• --scope user — across all your projects (stored in ~/.claude.json);
• --scope project — in the .mcp.json file at the project root, committed to git and working for the whole team.

To check what got connected, run /mcp inside a session or claude mcp list in the terminal.

Where a beginner should start: context7 gives the agent current library documentation (curing hallucinations about outdated APIs), and github gives it issues and pull requests. Cards for both servers, plus a full catalog with connection commands, are in the MCP servers section. And do read up on connecting MCP safely: a third-party server gets access to your environment, and trusting it blindly is a bad idea.

Which skills should a beginner install?

Skills are packaged sets of instructions for a specific job: code review, commit generation, an SEO audit. Plugins are the delivery format and skills are the content: you install plugins and you use skills. They're invoked with a namespaced slash command, for example /commit-commands:commit.

Installation goes through the marketplace:

/plugin install <name>@<marketplace>

The official marketplace is connected out of the box. Once installed, invoke the skill with its slash command — it does its job and stops. Start small: one or two skills for your recurring tasks, not everything at once, otherwise the command list balloons and you drown in it. An overview of proven options is in the Claude skills section.

How do you avoid spending a week on setup?

The honest truth: installing Claude Code takes five minutes. Getting it to the point where it reliably does what you need takes weeks of trial and error. You have to write a solid CLAUDE.md, pick and wire up MCP servers, configure subagents, skills and commands for your stack, and not break your security along the way. For a beginner that's the most expensive part in terms of time.

That assembly work is exactly what our Quest engine handles — a ready-made set of rules, agents, MCP wiring and skills for Claude Code and Cursor, put together so the agent works predictably from day one. You can build a configuration for your project in the studio or see what's inside the engine. It isn't a replacement for this guide — more a way to skip a month of learning the hard way. And the basic commands and MCP are still worth knowing: the engine uses them too.

FAQ

Do I need a paid subscription for Claude Code?

Yes. It works with a Claude Pro, Max, Team or Enterprise subscription, or with an Anthropic API key (pay per token). The free Claude.ai tier doesn't include Claude Code.

Should I install via npm or the native installer?

As of July 2026 the native installer (install.sh / install.ps1) is recommended — it doesn't require Node.js. The npm package is also supported and installs the same binary under the hood; pick it if you already have Node.js 22+ set up.

Do I have to connect MCP to get started?

No. The plain CLI is enough for your first tasks: the agent already sees your files, edits code and runs commands. MCP comes later, when you need current documentation or access to a database or a browser.

What do I do when the agent does the wrong thing?

First run /clear to drop the accumulated context. Then sharpen how you phrased the task and write the project's rules into CLAUDE.md — most "misunderstandings" are cured by that, not by switching models.

How do I know the context is running out?

The /context command shows how much of the context window is in use. When the number gets close to the limit, run /compact — the CLI compresses the conversation history while keeping the substance, and you carry on without losing the thread.