🟠 Claude Code: Anthropic's coding agent in your terminal

What Claude Code actually is

Claude Code is Anthropic’s terminal-native coding agent. You run claude inside a project; it reads files, edits them, runs your tests and shell commands, and commits — after asking your permission for the risky parts.

It is not a chat window with a copy-paste workflow. The agent loop is the product: it looks at the repo, decides a next action, executes a tool, reads the result, and repeats until the task is done. What an agent loop is →

Three things worth knowing before installing:

• It runs on your machine, in your terminal, against your real files.
• It needs a paid Claude plan or API access — there is no permanent free tier.
• The “intelligence” is in the cloud; the hands are local. That split is what makes it useful and what makes permissions matter.

Step 1 — Install

Pick one method. The native installer is the recommended one and auto-updates in the background.

macOS / Linux / WSL:

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

Windows (PowerShell):

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

Windows (WinGet):

winget install Anthropic.ClaudeCode

macOS (Homebrew) — note: Homebrew installs do not auto-update:

brew install --cask claude-code

Any OS (npm) — requires Node.js 22+:

npm install -g @anthropic-ai/claude-code

Verify it landed:

claude --version
# 2.1.319 (Claude Code)

Step 2 — Sign in

cd ~/projects/my-app
claude

On first launch it asks how to authenticate. Pick Claude account and finish the browser flow. You need one of:

On first login with a Console account, Claude Code creates a dedicated workspace so you can track what the agent spends.

Step 3 — Your first session

Inside the session, talk to it like a colleague who can see the repo:

> explain the architecture of this project, then list the three riskiest files

Then give it something real:

> the login form shows a blank screen after a wrong password — find the cause and fix it

Claude Code will read files, propose edits as diffs, and ask before running commands. Read the diffs. Approving blindly is how people ship broken code with an agent’s name on the commit.

The commands you’ll actually use

Step 4 — Teach it your project (CLAUDE.md)

/init creates a CLAUDE.md at the repo root. This file is injected into every session — it’s how you stop repeating yourself:

# Project conventions
- Python 3.12, FastAPI, tests with pytest
- Run `pytest -q` before claiming a task is done
- Never touch `migrations/` without asking
- French UI strings, English code comments

Keep it short and factual. A 200-line CLAUDE.md eats context on every single turn. Why context is a budget →

Step 5 — Headless mode

The same agent works as a Unix command, which is where it becomes infrastructure:

claude -p "summarize what changed in the last 5 commits" 
claude -p "list every TODO in src/ as a markdown checklist" > todos.md

-p (print mode) runs one turn, prints the answer, exits. Perfect for CI, git hooks, cron jobs, or piping into another tool.

The permission model (don’t skip this)

Claude Code asks before actions with side effects. Three outcomes per request: allow once, allow for this session, always allow (written to settings).

• Start restrictive. Widen only for commands you’d let a junior run unattended.
• Never blanket-allow rm, git push, or anything touching production credentials.
• Agent security is a real field now: prompt injection through files and web pages is the main attack path. Agentic risks →

Claude Code vs. the alternatives

The pragmatic answer for most people: try Gemini CLI free, keep Claude Code if you already pay for Claude, and use Codex CLI if you want the same terminal workflow wired to a local model.