Environment Setup

Get one AI coding agent installed and verified. Fifteen minutes, done once, never thought about again.

The hook

Every lab in this course ends the same way: your agent does something real while you watch. And none of it works if the very first step of your journey is a blinking command not found. So we clear that hurdle now — install once, verify once, and then put it out of your mind for good.

What you need (plain English)

Three things, and the order matters:

1. A terminal you're comfortable in — a macOS or Linux shell, or Windows with WSL.
2. One agent installed — Claude Code or OpenCode is plenty for every lesson here. Codex and Grok are optional extras, shown only where their mechanics diverge.
3. A throwaway practice repo — loops learn by doing, and you will badly want a sandbox where a mistake costs you exactly nothing.

[!WARNING] Mechanics change weekly. Treat the commands below as the shape of installation, not scripture. The authoritative steps always live in each tool's docs — linked in every section. If a command here ever disagrees with the official docs, the docs win, no contest.

Install and verify

# Claude Code — live docs: https://docs.claude.com/en/docs/claude-code
npm install -g @anthropic-ai/claude-code
claude --version
claude   # first run walks you through login

# OpenCode — live docs: https://opencode.ai/docs
curl -fsSL https://opencode.ai/install | bash
opencode --version
opencode auth login

Want the optional extras? They follow the identical rhythm — install, --version, authenticate: Codex (see OpenAI's live docs) and Grok CLI (see xAI's live docs).

Make your sandbox

mkdir loop-sandbox && cd loop-sandbox && git init
claude "create a tiny node project with one failing test"

mkdir loop-sandbox && cd loop-sandbox && git init
opencode run "create a tiny node project with one failing test"

That failing test is no accident — you planted it on purpose. It's the raw material your first loop will chew on in Project 1.

[!NOTE] Going deeper: why insist on a throwaway repo? Because your week-1 loops run at L1 (report-only) for exactly this reason — so a wrong move is cheap. It's the same discipline this repo applied to its own loops, which all started report-only. The full permission ladder (L1→L3) is laid out in the agentic coding primer.

Check yourself

Q: The install command on this page fails with an error you've never seen. What's the course-approved first move?

Answer

Open the tool's live documentation — the link is right above — not a search engine, and not this page. Mechanics change weekly; this course deliberately teaches lasting shapes and treats every command as a pointer to the current docs.

Try With AI

The moment one agent works, ask it:

"Check this machine for everything a second AI coding agent would need — package manager, node/python versions, auth files — and report what's missing. Don't install anything."

Notice what just happened: you ran your first report-only (L1) task — the exact mode every loop in this course is born in.

When it goes wrong

---

Glossary terms used on this page: agent, L1 report-only, sandbox — see ../02-foundations/glossary.md.

Sources: this setup guide follows Panaversity's Agentic Coding Crash Course (S2); commands are pointers to each tool's live documentation (attribution policy rule 3). Full attribution: resources/sources.md.