Claude Code is Anthropic's terminal-first agent. Unlike a chat that pastes diffs at you, it operates inside your project, reading files, editing them in place, running commands, and asking before doing anything destructive. The headline is not a feature; it is that the friction of using an agent has dropped to nearly zero.
It is the tool to reach for when the change touches more than a single file, or when plan-before-edit discipline is what the work needs. The fact that it ships with hooks, skills, MCP support, and slash commands out of the box means almost everything else in this directory plugs into it.
§01What you actually get
The agent reads, writes, edits, and searches the codebase within a sandboxed scope. Plan mode separates "decide" from "do," which is the single biggest behavioral change versus chat-style coding. Anthropic Skills and Model Context Protocol servers are first-class citizens, not bolt-ons, which means a community ecosystem of integrations grew up around them faster than around competing tools.
Lifecycle hooks fire on tool use, session events, and errors, and they are the basis for most of the community mods you find on this site. Subagents let the main loop spawn focused agents for specific subtasks (research, code review, doc generation), so you can keep the parent context clean.
§02When it earns its keep
The sweet spot is multi-file changes in a codebase you have worked in before. The agent can hold context that a chat-style tool cannot, and the friction of giving it more context is small: a CLAUDE.md, a few skills, an MCP server or two. For one-shot scripts it is overkill. For green-field projects with no conventions yet, it is actually liberating, because the agent will invent reasonable defaults that you can later codify.
§03Setup
# install
npm install -g @anthropic-ai/claude-code
# log in
claude auth
# in your project root
cd ~/Code/your-project
claude
First-time setup takes ten seconds. The decision that takes longer is what to put in CLAUDE.md. Treat it as the system prompt your repo deserves.
◆ pull quote
“Claude Code is not a chatbot in your terminal. It is an agent in your repo. The mental model matters; you brief it the way you would brief a junior, not the way you would ask a search engine.”
§04Caveats
Tokens cost money, and an unsupervised agent in plan-execute loops can burn through them faster than expected; setting a budget is part of the setup, not optional. Conventions matter more than they do in chat tools: without a CLAUDE.md the agent invents conventions, and with one, it follows yours. Hooks run shell commands on lifecycle events, so the rule of thumb is to never install a hook you have not read.