Claude Code Setup Guide 2026: Install to First Shipped Feature in 30 Minutes

What Is Claude Code, in One Paragraph?

Claude Code is Anthropic's coding agent that runs in your terminal. Unlike a chat window you paste code into, it has direct access to your project: it reads files, edits them, runs commands, executes your tests, and iterates on failures until things pass. You describe outcomes; it does the work and shows you diffs. It's the tool I ranked #1 for developers in my vibe coding tools guide, and it's what I build client projects with. (Full opinions in my Claude Code review.)

What Do You Need Before Installing?

A terminal, a project (existing codebase or empty folder, both fine), and a paid account. As of mid-2026, Claude Code is included with:

• Claude Pro — $20/month. Enough for light-to-moderate daily use. Start here.
• Claude Max — $100/month (5x) or $200/month (20x). For people running it hours a day. You'll know when you're that person because Pro's limits will tell you.
• API billing — pay per use. My real-world client months land around $30–50. Good for variable workloads; less predictable than a flat plan.

The free Claude.ai plan does not include Claude Code. At $20/month for a tool that regularly saves me 10+ hours a month, the math isn't close — that's roughly $2/hour for senior-level output. I've spent more on worse.

Step 1: How Do You Install Claude Code?

Two routes as of mid-2026. The native installer is what Anthropic now recommends — no Node.js required, auto-updates in the background:

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

The npm route still works fine if you're already in Node-land:

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

Then move into a project and start it:

cd your-project
claude

That's the whole install. If something's broken it's almost always PATH or an ancient Node version — the native installer sidesteps both, which is why it's the default now.

Step 2: How Do You Log In?

First launch, type /login. It opens a browser, you authenticate with your Claude account (or Console account for API billing), and you're done. The credential persists across sessions. If you ever see auth weirdness after switching between plan and API billing, /logout then /login fixes 90% of it.

Step 3: Why Is CLAUDE.md the Highest-Leverage 15 Minutes?

Here's the part most setup guides skip, and it's the part that separates "neat demo" from "tool I trust with client work." CLAUDE.md is a markdown file at your project root that Claude Code reads at the start of every session. Without it, the agent infers your stack and conventions fresh every time — and infers means guesses.

Run /init and Claude Code will scan the project and draft one for you. Then edit it like documentation for a sharp new hire. Mine follow this shape:

• Stack and versions. "Rails 8, SQLite in production, Hotwire — no React." Saves you from watching an agent cheerfully add a webpack config.
• Commands. How to run tests, lint, and start the dev server. The agent uses these to verify its own work, which is the entire trick.
• Conventions with teeth. "One state management pattern. Services go in app/services. Never commit directly to main."
• Known gotchas. The stuff that burned you before. My files have a literal "learned the hard way" section, and it has paid for itself repeatedly.

Keep it under a couple hundred lines. A CLAUDE.md that's a novel gets skimmed — by the agent and by you.

Step 4: Which MCP Servers Should You Add?

MCP (Model Context Protocol) is the open standard that connects Claude Code to external tools — documentation, databases, browsers, project trackers. Adding one is a wizard:

claude mcp add

Pick from the list or paste a server command. My advice after a year of client setups: start with one or two, not twelve. A documentation server (so the agent pulls current library docs instead of trusting its training data) is the single highest-value add for most stacks. A browser or database server earns its place the first time you need it, not before. Every server you add is context the agent carries around; a wall of MCP servers mostly just feels productive.

If you're on a specialty stack, this is where it pays off fast — Flutter folks, see how the Dart MCP server changes the game in vibe coding Flutter.

Step 5: What Should Your First Project Be?

Not "build me an app." I know. But the fastest way to learn how Claude Code thinks is a task you can verify: a real bug on an existing project, a small feature plus its test, the boring script you've deferred for months. My standard first-session loop:

1. Ask it to explain before it touches anything. "Walk me through how auth works in this codebase." You learn whether it actually understands the project, and so does it.

2. One scoped task. "Add validation to the signup form email field and write a test for the rejection case." Concrete, checkable, small.

3. Read the diff. Claude Code shows you changes before committing anything. Read them. This habit is the difference between using an agent and being used by one.

4. Make it verify. "Run the tests." The agent fixing its own red tests is where the magic actually lives.

Graduate to bigger autonomous runs as trust builds. On client work I now hand it multi-file features whole — but I earned that trust one readable diff at a time, and so should you.

How Does Claude Code Compare to the Alternatives?

The short version: Cursor if you want a visual IDE and an easier first hour; Codex if you're OpenAI-committed; the full field is in best AI coding agents for 2026. Claude Code's edge is depth — codebase understanding, autonomous multi-step runs, and backend work. It's a power tool with a learning curve, not a toy with a tutorial.

What Are the Mistakes Every Beginner Makes?

Skipping CLAUDE.md. Then wondering why the agent keeps making the same wrong assumption. It's not psychic; it's literate. Write things down.

Prompting like a wish, not a spec. "Make it better" gets you a refactor you didn't want. "Extract the payment logic into a service object and keep the controller under 20 lines" gets you software.

Accepting diffs unread. The output looks confident because it's always confident. Confidence is not correctness — ask me how I know, or just read what happens to unreviewed AI code in production.

Shipping without a security pass. Claude Code writes better code than any other agent I've tested, and it will still happily implement the insecure thing you asked for. Before anything touches real users or real money, run the 20-point security checklist.

Your 30-Minute Setup, Recapped

Minutes 0–5: install (native installer or npm), claude, /login. Minutes 5–20: /init, then edit CLAUDE.md with your stack, commands, and conventions. Minutes 20–25: claude mcp add for one documentation server. Minutes 25–30: first scoped task, read the diff, make it run the tests. That's the whole setup — everything past this is reps.

Want to see what Claude Code looks like on real production work — not demos? Vibe Coding with Claude covers my full client workflow. And if you've already got an AI-built app that needs adult supervision before launch, book a free strategy call — I'll tell you honestly whether it's shippable or a rescue.