Most AI coding tools live inside an editor and autocomplete the line you are typing. Claude Code takes a different shape: it is a command-line tool that treats your whole project as the workspace. You describe a goal in plain language, and it reads the relevant files, makes edits across as many of them as the task needs, runs your tests and build, and iterates until the job is done — all from the terminal, with you watching and steering. Built by Anthropic and powered by Claude models, it is agentic by design rather than a fancier autocomplete.
This guide explains what Claude Code is, how to install and start it, what actually happens during a session, how its memory and customization work, where it runs beyond the terminal, and how access is billed. It assumes you can use a terminal; everything else it spells out.
What Claude Code is
Claude Code is a terminal-native agentic coding assistant. The key word is agentic: instead of answering once and stopping, it loops — gathering context, taking action, and verifying the result — keeping the full conversation in a session you can interrupt and redirect. It can read and edit files, run shell commands, search the web for documentation, and create commits, all under a permission model you control.
Installing and starting it
Installation is a one-line script on every major platform, and you launch it by running claude inside a project folder:
# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash
# Windows (PowerShell)
irm https://claude.ai/install.ps1 | iex
# then, inside your project
cd your-project
claudeThe first run signs you in and drops you into an interactive session. You can also skip the chat for quick jobs: claude "add a health check endpoint and a test" runs a single task, and claude -p "summarize today's git log" runs in print mode for scripting.
How a session actually works
Ask for something and Claude Code runs a loop. It first builds understanding — reading files, grepping for patterns, sometimes checking docs. Then it acts: editing files, running tests and commands. Then it verifies, re-running tests and adjusting if something failed. It repeats that cycle until the task is complete or it needs your input.
Crucially, it is conversational, not fire-and-forget. By default it asks before editing files or running commands; you can approve changes one by one, accept all for a session, or pre-approve trusted commands in .claude/settings.json. Press Esc to interrupt, or just type a correction while it works.
Memory: CLAUDE.md and what Claude learns
To avoid re-explaining your project every session, Claude Code reads a CLAUDE.md file from your project root (and ~/.claude/CLAUDE.md for personal, cross-project rules) at startup. Put your architecture notes, build and test commands, and coding conventions there and they become persistent context. Running /init generates a starting CLAUDE.md for an existing codebase.
Separately, Claude Code keeps an auto-memory of things it learns as it works — commands it discovered, fixes that worked — so it improves across sessions without you spelling everything out. One honest caveat: CLAUDE.md is context Claude tries to follow, not an enforced rule. For things that must hold every single time, use a hook (below) instead.
Making it your own
Claude Code is built to be extended, and the pieces differ by who is in control and when they load. CLAUDE.md is always-on context. Skills (invoked as slash commands like /deploy) are reusable workflows that load only when used. Hooks are shell commands that fire deterministically on events — perfect for "always format after editing." Subagents are isolated workers that handle a noisy task and return only a summary. And MCP servers connect Claude to external systems.
That last one is worth a beat. Through the Model Context Protocol, Claude Code can read from and act on tools it cannot reach natively — GitHub, Slack, Jira, a database — configured in your settings file. (If MCP is new to you, our separate MCP guide covers it end to end.)
Where it runs: terminal, IDEs, and CI
The terminal is home, but not the only surface. There are first-class extensions for VS Code and JetBrains IDEs with inline diffs and shared context, a desktop app, and a browser version at claude.ai/code. For automation, the same tool runs headless in CI — wiring Claude Code into GitHub Actions to review pull requests or triage issues is a common pattern.
How access and pricing work
Claude Code is not billed as a separate product — it runs on Claude models through whatever account you already have. You can use it on a Claude subscription (Pro, Max, or Team), where it draws on your plan's usage allowance, or on a pay-as-you-go API account billed per token, which suits automation and scale. It also works through Amazon Bedrock, Google Vertex AI, and Microsoft Foundry for teams standardized on a cloud provider. The practical takeaway: you are paying for the model usage Claude Code drives, not for the tool itself.
The fastest way to understand Claude Code is to point it at a real task in a real repo — fix a failing test, add a small endpoint, refactor a function — and watch the gather-act-verify loop work, correcting it as you go. A few sessions in, the CLAUDE.md file, a couple of custom skills, and one or two MCP connections turn it from a clever assistant into a teammate that already knows your project.