Install Node.js
Both agents install through Node, so this comes first, and yes, straight from the terminal. Mac (works on a brand new machine; the sudo line asks for your Mac password):
NODE_V=$(curl -fsSL https://nodejs.org/dist/index.tab 2>/dev/null | awk 'NR>1 && $10 != "-" {print $1; exit}')
curl -fL -o /tmp/node.pkg "https://nodejs.org/dist/$NODE_V/node-$NODE_V.pkg"
sudo installer -pkg /tmp/node.pkg -target /Linux (Debian and Ubuntu; other distros, use your package manager):
curl -fsSL https://deb.nodesource.com/setup_lts.x | sudo bash - sudo apt-get install -y nodejs
Windows (PowerShell):
winget install OpenJS.NodeJS.LTS
Prefer clicking? The LTS installer at nodejs.org does the same job. Either way, close and reopen your terminal when it finishes.
Install your agent
Open a terminal (Terminal on Mac, PowerShell on Windows). On Windows, drop the sudo from the front. Claude Code:
sudo npm install -g @anthropic-ai/claude-code
And Codex:
sudo npm install -g @openai/codex
One is plenty to start. Install both if you want both launch commands working at the end.
Make sure your terminal can find it
Sometimes the install finishes but the terminal still says command not found. The installer prints a fix you are supposed to copy out of a wall of text; skip that and run ours instead. It works immediately, no restart needed. If claude or codex already run, skip this step.
NPM_BIN="$(npm config get prefix)/bin"; PROFILE="$HOME/.zshrc"; [ -n "$BASH_VERSION" ] && PROFILE="$HOME/.bashrc"; grep -qs "$NPM_BIN" "$PROFILE" || echo "export PATH=\"$NPM_BIN:\$PATH\"" >> "$PROFILE"; export PATH="$NPM_BIN:$PATH"; echo "Done. $NPM_BIN is on your PATH (saved in $PROFILE)."
On Windows (PowerShell):
[Environment]::SetEnvironmentVariable("Path", "$([Environment]::GetEnvironmentVariable('Path','User'));$env:AppData\npm", "User"); $env:Path += ";$env:AppData\npm"; Write-Host "Done. The npm global folder is on your PATH."Launch it once
Type claude (or codex) in your terminal and sign in when it asks. That first sign-in is the only setup it needs.
Paste this prompt
One prompt, any system: it works out where it is running, creates your AI folder, adds the CLAUDE.md and AGENTS.md notes agents actually read, and wires up ai-claude and ai-codex so either agent launches in the right place, at full reasoning power, ready to work.
Set up my AI workspace. Here is what I want: 1. Figure out what operating system and shell you are running in, and tell me. 2. Create a folder called AI in the right place for this system: my home directory on Mac or Linux (~/AI), or C:\AI on Windows. Skip this if it already exists. 3. Inside that folder, create a CLAUDE.md and an AGENTS.md with the same short note: this is my AI workspace, where my agent projects, notes, and experiments live. Any agent launched here should treat it as home base. 4. Set up two launch commands in the right profile for my shell (zsh, bash, or PowerShell): - ai-claude: change into the AI folder and run claude --dangerously-skip-permissions --effort max - ai-codex: change into the AI folder and run codex --yolo -c model_reasoning_effort=xhigh Create both even if only one agent is installed, and tell me which of them will work right now. If either reasoning option is not recognized by the installed version, find the current way to set that agent to its highest reasoning level and use that instead. 5. Show me exactly what you added and where, then tell me to open a new terminal and try ai-claude or ai-codex. Explain each step in plain language as you go. If anything already exists, do not overwrite it, tell me instead.
Know what you just turned on
Agents normally ask before running commands or changing files. --dangerously-skip-permissions (Codex calls it --yolo) turns those confirmations off. That is what makes an agent feel like a coworker instead of a permission quiz, and it also means it acts without asking first. Keep anything precious backed up, let the agent live in your AI folder while you build trust, and any time you want the guardrails back, launch plain claude or codex.