Claude Code: The Complete Guide to Anthropic's AI Coding Agent

Claude Code is Anthropic's terminal-native AI coding agent. You describe a task — fix a bug, refactor a module, scaffold a new feature — and Claude Code plans the approach, edits files across your project, runs tests, and commits the result. It operates directly in your shell with full access to your codebase, your build tools, and your git history. Unlike IDE copilots that suggest one line at a time, Claude Code handles multi-step engineering tasks end to end. Launched in February 2025, it's used by engineering teams at Ramp, Shopify, Spotify, and thousands of individual developers. This guide covers everything: how it works, what it costs, how it compares to alternatives, and the practices that make it most effective.

What Is Claude Code?

Claude Code is a CLI application powered by Claude models with tool-use and extended thinking capabilities. You install it with a one-line native installer (or via Homebrew or WinGet), run claude in your terminal, and interact through natural language. It reads your project files, understands relationships between modules, and executes multi-step tasks using a loop of reasoning, tool calls, and validation.

The key distinction from IDE-embedded tools like Cursor or GitHub Copilot is architectural. Claude Code is an agentic system — it doesn't suggest edits for you to accept. It plans, acts, observes the result, and iterates. A prompt like "add rate limiting to the API endpoints with Redis backing" triggers file reads, dependency installation, code generation, test creation, and a structured commit. You review the output like a pull request from a colleague, not a suggestion from an autocomplete engine.

Claude Code runs on macOS 13+, Linux (Ubuntu 20.04+, Debian 10+, Alpine 3.19+), and Windows 10+ natively via PowerShell, CMD, or Git Bash. Windows also works through WSL. It works with any programming language and any project structure — monorepo, microservices, single-file scripts. If it lives in your filesystem and has a build command, Claude Code can work with it.

Getting Started

Installation takes one command:

# macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# Also available via Homebrew (macOS) or WinGet (Windows)
brew install --cask claude-code
winget install Anthropic.ClaudeCode

The native installer auto-updates in the background. (npm installation via npm install -g @anthropic-ai/claude-code still works but is deprecated.)

After installation, run claude in your project directory. Claude Code detects your project structure, reads any CLAUDE.md configuration files, and starts an interactive session. You'll authenticate with your Anthropic account on first run. Claude Code requires a Pro, Max, Team, Enterprise, or Console account — the free Claude.ai plan does not include access.

The single most impactful setup step is creating a CLAUDE.md file at your repository root. This file tells Claude Code how your project works — build commands, test commands, coding conventions, architectural decisions, and anything a new team member would need to know. Claude Code reads it automatically at session start, so every interaction respects your project's specific context. See our deep dive on the memory system for the full picture.

For first-time users: start with a small, well-tested task. Ask Claude Code to fix a failing test, add a missing type annotation, or refactor a function. Watch how it plans the approach, makes edits, and verifies the result. This builds intuition for how much context to provide and when to intervene. For step-by-step installation instructions, see How to Install Claude Code.

How Claude Code Works

Claude Code's architecture has three layers that build on each other.

Layer 1: Project Context (CLAUDE.md). Every session starts by reading CLAUDE.md files — project-level instructions that define your build commands, testing strategy, coding standards, and architectural constraints. This replaces the "explain your whole project every time" problem. Teams version-control these files alongside their code, so Claude Code's behavior stays consistent across the whole team. The memory system extends this with auto-generated memories that persist context across sessions — things you've taught Claude Code about your project accumulate over time.

Layer 2: Tool Connectivity (MCP). The Model Context Protocol is an open standard that connects Claude Code to external data sources and tools. Through MCP servers, Claude Code can query your database, check CI status, search documentation, or interact with any API that has an MCP adapter. Over 1,000 MCP integrations exist across the community. This layer turns Claude Code from a filesystem-only tool into one that understands your entire development environment. Our guide on MCP vs CLI vs Skills explains when each integration approach makes sense.

Layer 3: Subagents. For large tasks, Claude Code spawns parallel subagents that each work in isolated git worktrees. The main agent decomposes the problem, delegates subtasks, and merges results. Three built-in subagent types define the permission model: Explore (read-only codebase search, runs on Haiku for speed), Plan (analysis without edits), and General-purpose (full tool access). Subagents cannot spawn other subagents — the main agent coordinates all delegation. You can also define custom subagents with focused system prompts, specific tool access, and independent permissions. A multi-file refactoring that takes 30 minutes sequentially can finish significantly faster through parallel execution.

Key Features

Subagents — parallel subagents in isolated git worktrees, each with its own context window and permission scope. Mix model tiers (Haiku for Explore subagents, Opus for implementation) to optimize cost. Define custom subagents with focused system prompts via Markdown files in .claude/agents/.

Memory system — CLAUDE.md for deliberate project instructions, auto memory for emergent knowledge learned during sessions. Context compounds across conversations instead of resetting.

Skills and hooks — Skills are SKILL.md files stored in .claude/skills/ directories that become slash commands and can also be invoked automatically by Claude when relevant. Legacy .claude/commands/ files still work. Hooks are shell commands triggered on lifecycle events (PreToolUse, PostToolUse) that enforce deterministic guardrails. Together they make Claude Code programmable by your team.

Voice mode — speak coding instructions naturally. The same agent loop, same tools, different input method. Useful for accessibility, mobile workflows, and thinking-out-loud development.

Remote control — start a long-running task in your terminal, then monitor and steer from your phone or any browser. Available on all paid plans (Pro, Max, Team, Enterprise).

Code review — automated PR review that runs multiple specialized agents in parallel, looking for logic errors, security vulnerabilities, and regressions. Currently a research preview, available for Teams and Enterprise plans.

Security scanning — point Claude Code at a codebase and it hunts for vulnerabilities using full project context, not pattern matching. Currently a research preview.

Programmatic usage / Agent SDK — run Claude Code non-interactively with claude -p for CI/CD pipelines, batch operations, and scripted workflows. The Agent SDK provides Python and TypeScript packages for full programmatic control. (The -p flag was previously called "headless mode.")

/btw side chains — ask a quick question mid-task without interrupting Claude's current work.

Prompt stashing (Ctrl+S) — queue your next prompt while Claude is still executing the current one. Your instructions run automatically when it finishes.

Claude Code vs Alternatives

The AI coding tool landscape in 2026 has distinct categories. Claude Code occupies the terminal-native agent space — it runs in your existing shell, works with any editor, and executes autonomously.

Claude Code vs Cursor: Cursor embeds AI into a VS Code fork with inline editing, tab completion, and a composer for multi-file changes. Claude Code runs in terminal with full agent autonomy. Cursor is better for developers who want AI-enhanced editing within a GUI. Claude Code is better for developers who want autonomous task execution and work primarily in terminal. Many developers use both — Cursor for interactive editing, Claude Code for autonomous tasks.

Claude Code vs GitHub Copilot: Copilot focuses on inline suggestions and chat within VS Code or JetBrains. Its Copilot Workspace feature proposes multi-file plans but doesn't execute them autonomously. Claude Code's agent loop — plan, execute, validate, iterate — goes further for complex tasks.

Claude Code vs OpenAI Codex: Codex (the CLI agent, not the legacy model) takes a similar terminal-native approach. Both are agentic, both work from the command line. Key differences are model quality, extension architecture, and multi-agent capabilities.

Claude Code vs Windsurf: Windsurf (formerly Codeium) offers an IDE with AI flows. Like Cursor, it's GUI-native rather than terminal-native.

For detailed comparisons, see our Claude Code vs Cursor analysis and the VS Code integration guide.

Best Practices

Use TDD as your control mechanism. A failing test gives Claude Code an unambiguous target. Without tests, you get plausible code that may not match intent. The Red Green Refactor workflow is the single most effective pattern for reliable AI-assisted development.

Invest in CLAUDE.md and Skills. Fifteen minutes writing clear project instructions saves hours of repeated prompting. The Superpowers project demonstrates how well-crafted SKILL.md files in .claude/skills/ replace thousands of lines of ad-hoc prompting. Combine this with Obsidian as a knowledge base for even richer context.

Choose the right extension point. Before building an MCP server, check if a SKILL.md file solves your problem. MCP vs CLI vs Skills provides the decision framework. Reserve MCP for genuine external integrations — databases, APIs, CI systems.

Set up harnesses for long tasks. Multi-hour agent sessions need structure: initializer agents, progress files, incremental checkpoints. Our guide to effective harnesses for long-running agents covers the patterns that work.

Decompose by file boundaries for subagents. When using parallel subagents, assign different files to different agents to avoid merge conflicts. Subagents details the decomposition strategies.

Pricing

Claude Code has no separate license fee. You pay for the underlying Claude model usage through one of several paths:

Subscription plans include Claude Code access with usage caps:

Claude Pro ($20/month) — includes limited Claude Code usage, good for evaluation
Claude Max ($100/month or $200/month tiers) — significantly higher usage caps and priority capacity for daily use
Claude Team ($25/seat/month) — team-level access with admin controls and Code Review
Claude Enterprise — custom pricing with SSO, advanced security, and admin governance

API billing — pay per input and output token via the Anthropic Console. Rates vary by model: Opus (most capable, highest cost), Sonnet (balanced), Haiku (fastest, cheapest). Average cost is roughly $6/developer/day according to Anthropic, with 90% of users spending under $12/day.

Start with Pro to evaluate, upgrade to Max if you hit rate limits. On API billing, route simple tasks to Sonnet or Haiku and reserve Opus for complex reasoning — this cuts costs significantly. For the full pricing breakdown, see How Much Does Claude Code Cost?.

All Claude Code Resources

Deep Dives

Features

Comparisons

FAQ

Glossary

Claude Code — Anthropic's terminal-based AI coding agent
Anthropic — AI safety company building Claude
Claude — Anthropic's family of large language models
Model Context Protocol (MCP) — open protocol for AI tool connectivity
Agentic Coding — AI-driven autonomous software development
Multi-Agent Systems — architectures where multiple AI agents collaborate
Agent Teams — coordinated sub-agents working in parallel
CLAUDE.md — project-level configuration for Claude Code

Topic Hub

Claude Code — Everything You Need to Know

Found this useful? Subscribe to AI News for daily AI briefings.