What Is Claude Code? The Complete Guide for 2026

Claude Code is Anthropic's AI coding agent. Per the official docs, it is "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Available in your terminal, IDE, desktop app, and browser." If you have heard the name on engineering Twitter and wondered whether to install it, this is the guide.

This post is for developers who have not used Claude Code before. We will go from zero to a working first session, cover the primitives that actually matter (CLAUDE.md, skills, hooks, MCP, subagents, plan mode), compare it honestly to the alternatives, and finish with five things to do after you install it. Every feature, command, and pricing number in this post is pulled from Anthropic's official documentation. Sources are listed at the end.

The 30-second primer

Claude Code is an AI pair programmer that runs across four surfaces: a CLI in your terminal, a VS Code extension, a JetBrains plugin, and a native desktop app for macOS and Windows. There is also a browser version at claude.ai/code. All surfaces "connect to the same underlying Claude Code engine, so your CLAUDE.md files, settings, and MCP servers work across all of them."

The important distinction: the terminal CLI is not an IDE plugin. It is a long-running agent process. You start it in a project directory with claude, describe what you want, and the agent reads, writes, edits, runs tests, and makes commits. Think of it as a teammate that can type into your shell, with the context you give it in a file called CLAUDE.md.

How it actually works

Claude Code runs an agent loop. You send a message, the model decides which tool to call, the harness executes it, and the result feeds back into the model. It keeps going until the task is done or it needs you to answer a question.

At the start of every session, the agent loads memory. Anthropic's docs describe two complementary memory systems:

• CLAUDE.md files: markdown you write with persistent instructions (coding standards, architecture, commands).
• Auto memory: notes Claude writes itself about things like "build commands, debugging insights, architecture notes, code style preferences, and workflow habits." Auto memory requires Claude Code v2.1.59 or later and lives under ~/.claude/projects/<project>/memory/.

CLAUDE.md files are loaded from a hierarchy: managed policy, project (./CLAUDE.md or ./.claude/CLAUDE.md), user (~/.claude/CLAUDE.md), and local (./CLAUDE.local.md). The docs recommend keeping each file "under 200 lines" because longer files "consume more context and reduce adherence."

Auto memory loads the first 200 lines or 25KB of MEMORY.md into every session. Topic files like debugging.md or patterns.md are not loaded at startup; Claude reads them on demand.

Install and first session

Per the official setup page, Claude Code runs on:

• macOS 13.0+
• Windows 10 1809+ or Windows Server 2019+
• Ubuntu 20.04+, Debian 10+, Alpine Linux 3.19+
• 4 GB+ RAM, x64 or ARM64

The native installer command from the docs:

curl -fsSL https://claude.ai/install.sh | bash

For Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

Homebrew and WinGet are also supported:

brew install --cask claude-code

winget install Anthropic.ClaudeCode

Native installations "automatically update in the background." Homebrew and WinGet do not.

After install, verify with claude --version, then go to any project and run claude. First launch opens a browser to log in. Anthropic's docs note that Claude Code "requires a Pro, Max, Team, Enterprise, or Console account. The free Claude.ai plan does not include Claude Code access." You can also authenticate through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry.

Core features in 2026

Claude Code in 2026 is not just "a CLI with a chat loop." It has an ecosystem of primitives. Here are the ones documented on code.claude.com.

CLAUDE.md and project memory

The single most important file in any Claude Code project is CLAUDE.md. Run /init inside a project and Anthropic's docs say Claude "analyzes your codebase and creates a file with build commands, test instructions, and project conventions it discovers."

CLAUDE.md files can import other files with @path/to/import syntax, with a maximum import depth of five hops. For large projects, you can scope rules to specific file paths with .claude/rules/ files using YAML frontmatter with a paths glob. This is the official way to keep project memory organized without bloating context.

Skills

Skills are the 2026 replacement for custom slash commands. Per the docs: "Custom commands have been merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way."

A skill is a directory with a SKILL.md file containing YAML frontmatter and markdown instructions. Frontmatter fields documented in the skills reference include name, description, disable-model-invocation, user-invocable, allowed-tools, model, effort, context, agent, hooks, paths, and shell.

Example from the docs:

---
name: explain-code
description: Explains code with visual diagrams and analogies. Use when explaining how code works.
---

When explaining code, always include:

1. **Start with an analogy**
2. **Draw a diagram** using ASCII art
3. **Walk through the code** step-by-step
4. **Highlight a gotcha**

Claude Code also ships with bundled skills like /simplify, /batch, /debug, /loop, and /claude-api. Skills live at four scopes in priority order: enterprise, personal (~/.claude/skills/), project (.claude/skills/), and plugin.

Hooks

Hooks are shell commands, HTTP calls, prompts, or subagents that run at specific points in Claude's lifecycle. The docs describe four hook types:

1. Command hooks (type: "command") run shell commands
2. HTTP hooks (type: "http") POST JSON to a URL
3. Prompt hooks (type: "prompt") ask a model for a yes/no decision
4. Agent hooks (type: "agent") spawn a subagent to validate

Documented lifecycle events include SessionStart, SessionEnd, UserPromptSubmit, PreToolUse, PostToolUse, Stop, Notification, SubagentStart, SubagentStop, PreCompact, PostCompact, WorktreeCreate, and more.

A typical use: run Prettier after every file edit, or block destructive rm -rf commands before they run. Hooks are configured in settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "./scripts/validate.sh"
          }
        ]
      }
    ]
  }
}

Model Context Protocol (MCP)

MCP is "an open standard for connecting AI tools to external data sources." In Claude Code, you add servers with claude mcp add. The docs show three installation scopes:

Add a server with a single command (examples from the docs):

claude mcp add --transport http stripe https://mcp.stripe.com
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

Then use /mcp inside a Claude Code session to authenticate and manage connections. MCP supports stdio, SSE, HTTP, and WebSocket transports, with OAuth 2.0 for remote servers.

Subagents

Subagents let a task run in its own isolated context window. Claude Code ships with built-in subagents including Explore, Plan, and general-purpose. Per the docs:

• Explore: "A fast, read-only agent optimized for searching and analyzing codebases." Uses Haiku for speed.
• Plan: "A research agent used during plan mode to gather context before presenting a plan." Read-only.
• General-purpose: "A capable agent for complex, multi-step tasks that require both exploration and action."

You can also create your own. Subagents are markdown files with YAML frontmatter, stored in .claude/agents/ (project) or ~/.claude/agents/ (user). Documented frontmatter fields include name, description, tools, disallowedTools, model, permissionMode, maxTurns, skills, mcpServers, hooks, memory, effort, background, isolation, and color.

Use /agents inside Claude Code to manage them interactively.

Plan mode

Plan mode is a read-only planning pass before Claude makes changes. Per the docs: "Plan Mode instructs Claude to create a plan by analyzing the codebase with read-only operations, perfect for exploring codebases, planning complex changes, or reviewing code safely."

Three ways to enter plan mode:

1. Press Shift+Tab twice during a session. You will see ⏸ plan mode on at the bottom.
2. Start a session with claude --permission-mode plan.
3. Run a headless query with claude --permission-mode plan -p "Analyze the auth system and suggest improvements".

Press Ctrl+G to open the proposed plan in your text editor for direct editing.

Agent teams and the Agent SDK

Subagents coordinate within a single session. For cross-session coordination, Anthropic documents agent teams. For fully custom agents built on Claude Code's internals, there is the Agent SDK, which the docs describe as letting you "build your own agents powered by Claude Code's tools and capabilities, with full control over orchestration, tool access, and permissions."

How Claude Code compares

The AI coding tool space in 2026 is crowded. Honest comparisons based on what each tool's docs actually say:

Claude Code vs Cursor

Cursor is an IDE (a fork of VS Code). Claude Code is primarily a CLI, with an IDE extension as an alternative surface. If your workflow is IDE-first and you want deep editor integration (inline tab completion, visual diffs on every keystroke, @-mentions), Cursor is built for that. If you live in the terminal and want to pipe, script, and chain coding tasks like Unix utilities ("Pipe logs into it, run it in CI, or chain it with other tools," per the Claude Code docs), Claude Code is built for that.

Claude Code also ships a VS Code extension that "provides inline diffs, @-mentions, plan review, and conversation history directly in your editor," so you do not have to choose one or the other.

Claude Code vs Codex CLI

OpenAI's Codex CLI is the closest shape match: also a CLI, also agentic, also runs on your machine. The differences are model lineage and the ecosystem around the CLI. Claude Code's docs document plugins, skills, hooks, MCP, subagents, agent teams, GitHub Actions integration, GitLab CI/CD integration, scheduled routines, and the Agent SDK as first-party features. Pick based on model preference and which ecosystem you want to build on.

Claude Code vs Aider

Aider is an open-source terminal coding assistant with strong git integration. Similar surface, different philosophy. Aider has been around longer and has a simpler, more opinionated feature set. Claude Code's docs describe a broader surface area: four install targets, IDE plugins, a web interface, Bedrock/Vertex/Foundry support, and the plugin ecosystem. Aider is lean and focused, Claude Code is a platform.

Pricing

Pulled from claude.com/pricing:

• Free: $0. Does not include Claude Code.
• Pro: $17/month annual or $20/month monthly. Claude Code is included.
• Max: From $100/month. Pro features plus "5x or 20x more usage than Pro," higher output limits, early access, and priority during high traffic.
• Team: $20/seat/month annual or $25/seat/month monthly (standard seat). Premium seat is $100/month annual or $125/month monthly. Includes Claude Code, SSO, admin controls.
• Enterprise: Seat price ($20) plus usage at API rates. Adds SCIM, audit logs, role-based access, compliance API, custom data retention, IP allowlisting, and a HIPAA-ready offering.

The setup docs confirm: "Claude Code requires a Pro, Max, Team, Enterprise, or Console account. The free Claude.ai plan does not include Claude Code access."

Five things to do after you install it

Based on the official docs, here is a sensible onboarding sequence.

1. Run /init to generate CLAUDE.md

The docs say /init "analyzes your codebase and creates a file with build commands, test instructions, and project conventions it discovers." Refine from there. If you want a structured template, we built a CLAUDE.md generator that walks you through the sections your project needs.

2. Try plan mode before shipping anything

Press Shift+Tab twice at the start of any non-trivial task. The docs are explicit: "A two-phase approach with planning produces better results than jumping straight to code." Plan mode uses read-only tools, gathers requirements, and proposes a plan you approve before execution.

3. Connect an MCP server

Pick one external tool you use every day (GitHub, Sentry, Linear, Stripe, a database) and connect it with claude mcp add. The difference between an agent that can only touch your filesystem and one that can read your production error logs is large. Browse the MCP Directory for curated servers.

4. Save a workflow as a skill

The moment you catch yourself pasting the same 10-line playbook into chat twice, convert it to a skill. Either create the directory manually at .claude/skills/<name>/SKILL.md or use our Skill Builder to scaffold one from a short description. Skills you want to keep in your back pocket can live in the Skills Directory.

5. Configure at least one hook

Start with something small. A PostToolUse hook that runs Prettier or ESLint after edits. A PreToolUse hook that blocks writes to production.env. A SessionEnd hook that logs session duration. The muscle memory of writing hooks unlocks the whole extensibility model.

Where it falls short

Being honest about limitations, based on what the docs say and do not say.

Cost for heavy users. Pro at $20/month includes usage caps. Heavy users hit Max at $100/month or more. For teams using it as a primary coding tool, costs scale with seat count and usage.

Context is finite. Even with auto memory, CLAUDE.md, and subagents to preserve main-conversation context, you will hit limits on large codebases. The docs mention /compact and re-injection patterns, but long sessions still require care.

Terminal-first bias. The CLI is the most feature-complete surface. The VS Code extension, JetBrains plugin, desktop app, and web version all exist, but the CLI gets features first. If your whole team lives in JetBrains, you may feel one step behind.

Plan mode does not prevent bad plans. It prevents file changes during planning, but if the plan is wrong, you are still responsible for catching it. The docs say "a two-phase approach produces better results" and that is true on average, but not a free pass to stop reading what the agent proposes.

Non-Anthropic models are not drop-in. You can run Claude Code on Bedrock, Vertex, or Foundry, but those are alternate Claude deployments, not arbitrary models. If you want to run Claude Code against GPT, Gemini, or an open model, that is not what the docs describe.

Start here

If you got this far and have not installed it, here is the install command one more time:

curl -fsSL https://claude.ai/install.sh | bash

Then cd into a project and run claude. Let /init write a first pass of CLAUDE.md. Ask it to explain your codebase. Watch the agent loop run.

If you want to go deeper:

• The CLAUDE.md Generator scaffolds a project memory file from a short description.
• The MCP Directory catalogs MCP servers you can plug in.
• The Skills Directory collects shareable skills.
• The Skill Builder turns a one-line prompt into a SKILL.md you can drop into .claude/skills/.

The fastest way to understand Claude Code is to give it a real task you would otherwise do yourself and watch what it does. Start there.

References

Every claim in this post is sourced from the following official documentation pages, fetched on 2026-04-19:

• code.claude.com/docs/en/overview: product description, surfaces (terminal, VS Code, JetBrains, Desktop, Web), and capabilities (commits/PRs, MCP, skills, hooks, subagents, schedules, Unix piping).
• code.claude.com/docs/en/setup: system requirements, install commands for macOS/Linux/Windows/Homebrew/WinGet, authentication requirements, npm install option, and auto-update behavior.
• code.claude.com/docs/en/memory: CLAUDE.md file locations, precedence, import syntax, .claude/rules/ scoping, and auto memory (v2.1.59+, 200-line/25KB load limit, storage location).
• code.claude.com/docs/en/hooks: hook types (command, http, prompt, agent), lifecycle events, configuration structure, matcher patterns.
• code.claude.com/docs/en/skills: SKILL.md structure, frontmatter fields, bundled skills, skill scopes, merging of custom commands into skills.
• code.claude.com/docs/en/mcp: claude mcp add commands, installation scopes (local/project/user), OAuth flow, transport types.
• code.claude.com/docs/en/subagents: built-in subagents (Explore, Plan, general-purpose), custom subagent frontmatter fields, /agents command.
• code.claude.com/docs/en/common-workflows: plan mode details (Shift+Tab cycle, --permission-mode plan, Ctrl+G to edit plans), worktrees, session resume.
• claude.com/pricing: Free/Pro/Max/Team/Enterprise pricing and feature breakdown.