What Is Claude Code?

Claude Code is Anthropic's official terminal-native coding agent. You run claude inside any project directory and it can read files, run shell commands, edit code, run tests, and explain its reasoning — all backed by Claude Sonnet 4 (or Opus / Haiku, your choice). Think of it as the missing CLI between you and the model.

What Claude Code can do

• Read any file in the repo — it indexes your project on first run.
• Edit files in place — applies diffs, asks before overwriting.
• Run shell commands — npm test, git diff, grep, etc.
• Run tests in a loop — fix until green.
• Use external tools (MCP servers) — connect to your DB, your linter, your custom CLI.
• Multi-file refactors — long-context model holds the whole module set.
• Interactive mode + headless (claude -p) mode — chat or one-shot.

How it differs from Cursor, Copilot, Aider

Most people end up with two: an inline editor (Cursor or Copilot) for while-you-type completions, and a terminal agent (Claude Code) for long tasks like "make all 14 tests pass" or "migrate this module to TypeScript."

How it bills

Two pricing surfaces:

• Anthropic Pro / Max subscription — flat $20–200/month, daily caps, tied to your Anthropic login.
• API key (pay-as-you-go) — pay only the tokens you actually consume. Required ANTHROPIC_API_KEY env var.

Cost comparison: Claude Code Proxy vs Official Subscription.

Install in 2 commands

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

cd your-project
claude

First run asks for your API key (or signs you into Anthropic Pro). After that, everything is one claude command away. Full guide: Claude Code Setup.

What a typical session looks like

$ cd my-app
$ claude

claude> the user signup endpoint is throwing 500. Find and fix.

[claude reads server/auth.ts]
[claude reads server/db.ts]
[claude runs: npm test --grep signup]
…
claude> The bcrypt comparison was using `===` instead of `bcrypt.compare`.
        I edited server/auth.ts (lines 47–52) and the test now passes.
        Want me to commit?

claude> yes

[claude runs: git add . && git commit -m "fix(auth): use bcrypt.compare"]

Tips for using it well

• Keep prompts specific. "Fix the bug in /api/users" beats "fix bugs."
• Let it run shell commands. Auto-approve npm test and git status — most of its value is in the loop.
• Use claude -p for one-shot scripts. Great in CI / makefiles.
• Watch token spend — multi-file edits in long projects can burn tokens fast. Sonnet is the sweet spot; Opus only when you need it.
• Add a CLAUDE.md to your repo — Claude reads it as a project manual on every session.

When NOT to use Claude Code

• Tiny ad-hoc scripts where the IDE plugin is faster (tab, done).
• Air-gapped environments where you can't reach the API.
• Code that requires GUI inspection (Cursor's preview is better for UI work).