How to Use Codex to Set Up a New Computer (No Coding Needed)

August 18, 2026: Wharton professor Ethan Mollick posted a story that's a better pitch for AI agents than most product demos: his father got a new MacBook after years without a laptop, didn't know how to set it up, and — on Mollick's advice — installed OpenAI's Codex CLI and simply asked it to take over. Codex solved his setup problems and installed what he needed. Mollick's line: "Family IT support will never be the same."

That's a different story from explainx.ai's earlier coverage of Mollick and Codex, where he had Codex scrape 5,032 of his own X bookmarks unsupervised for two hours. This one is smaller, more relatable, and more useful to most readers — it's the first real-world proof that a terminal AI agent built for developers can double as a computer setup assistant for someone who has never opened a terminal in their life. Here's how to actually do this, and the guardrails that keep it from becoming a cautionary tale instead.

TL;DR

What Codex actually is, in plain terms

Codex is OpenAI's terminal-based coding agent — you run it from a command-line app called Terminal (macOS/Linux) or PowerShell (Windows), and it reads and writes files, runs commands, and installs software on your behalf, one step at a time. It was built for developers, but nothing about it requires writing code yourself — you talk to it in plain English, and it decides which terminal commands to run to do what you asked.

That's exactly the gap Mollick's father's story closes. Setting up a new machine is mostly repetitive terminal-shaped work — installing package managers, downloading and configuring apps, adjusting system defaults — that a non-technical person finds intimidating specifically because it lives in a black terminal window with a blinking cursor. Codex removes the part that's actually hard (knowing which command does what) and leaves the part anyone can do (describing what they want in English).

Installing Codex CLI

After installing, open Terminal (search for it in Spotlight on macOS, or use PowerShell on Windows) and type codex. It will prompt you to sign in with a ChatGPT account — a Plus or Pro subscription unlocks higher usage limits, but a free account works for light use.

For someone doing this for the first time: installing Codex itself is the one step that still requires a small amount of hand-holding — either a family member walks them through the one-line install command over a call, or does it remotely once, before handing the terminal back with Codex already running.

The safety setting that matters most

This is the part most viral "I let an AI set up my computer" stories skip, and it's the one that actually determines whether the story ends well. Codex CLI has two separate settings that control how much it can do without asking first:

• sandbox_mode — what Codex is allowed to touch: read-only, workspace-write, or danger-full-access.
• approval_policy — when it has to ask before acting: on-request, untrusted, or never.

Don't run danger-full-access with never unattended on a primary machine. OpenAI's own Codex lead confirmed a July 2026 incident where GPT-5.6, running in full access mode with sandboxing and auto-review both disabled, tried to redirect a temp-file operation and instead deleted a user's entire $HOME directory. The failure chain was specific — full access removed the safety layers, the model tried to work around a path issue, and an honest mistake became a destructive one with nothing left to catch it.

For a computer-setup task, the safer combination is --sandbox workspace-write --ask-for-approval on-request (or the short form -s workspace-write -a on-request). Codex can install software and make changes, but it pauses and asks in plain English before anything it flags as higher-risk — which is exactly the moment a first-time user should be looking at the screen anyway. Note that as of CLI v0.147.0 (August 7, 2026), the older codex exec --full-auto shortcut has been removed — use the explicit sandbox and approval flags above instead.

What to actually ask it

An open-ended "set up my whole computer" prompt works, but a first-timer gets a better experience — and a safer one, since each step is easier to sanity-check — from specific, scoped requests:

• "Install Chrome, Spotify, and Zoom, and set Chrome as my default browser."
• "Turn on iCloud Drive, set up Photos to sync from my old phone, and enable Find My Mac."
• "Clean up the default Dock — remove apps I'm not going to use and add Chrome and Mail."
• "Set up my printer and check that it shows up when I try to print a test page."
• "Organize my Downloads folder by file type so it isn't one giant pile."

Each of these is small enough that Codex's step-by-step, ask-before-acting behavior stays legible — you can watch it reason through each command instead of trusting a black box for an hour-long unattended run.

When you want more than one-off tasks

Codex CLI on its own is a single assistant in a single terminal window — great for a one-time setup, but it doesn't remember across sessions or coordinate multiple tasks running in parallel. For readers who want to go further — running several agents at once, each with a role, coordinated by one you talk to — open-source projects have started building multi-agent "harnesses" on top of tools like Codex.

Munder Difflin is one example: a free, MIT-licensed desktop app that wraps Codex alongside Claude Code, Gemini's Antigravity, Grok, and several other agent CLIs, gives each one memory and a mailbox, and visualizes them as characters working a shared 2D office floor — coordinated by one orchestrator agent you talk to. It's a prototype (v0.4.4 as of this writing, roughly 1.8K GitHub stars), not something to hand a first-time laptop user, but it's a useful next step for anyone who reads Mollick's story and wants to know what "one Codex session" scales up into. It's also a concrete example of the pattern explainx.ai covers in loop engineering and MCP — a supervising loop routing work between tool-connected agents — which, together with Agent Skills, are the three skills explainx.ai flagged as mattering most heading into 2027.

What people are asking

"Isn't this just Copilot or ChatGPT with extra steps?" No — the difference is that Codex acts, not just answers. Asking ChatGPT "how do I install Chrome" gets you instructions to follow yourself; asking Codex CLI the same thing gets the install actually run, with Codex reading any errors and adjusting. That's the same agentic-vs-chat distinction explainx.ai covers in the AI skills developer roadmap's agentic-coding-tools stage — it applies just as much to a non-developer's first laptop as it does to a professional codebase.

"What if Codex gets something wrong?" With on-request approval, it asks before higher-risk actions, and every command it runs is visible in the terminal as it happens — nothing is hidden. If something looks wrong, you can say no or stop it mid-task. That visibility is exactly what full-access, unattended mode removes, which is why the safety section above matters more than any single feature.

"Does this work on Windows and Linux too?" Yes. Codex CLI runs natively on Windows via PowerShell (with a Windows sandbox) or under WSL2, and on Linux directly. The setup story here uses macOS as the example because that's Mollick's father's case, but the install commands and safety settings are the same across platforms.

"My relative is even less technical than 'can use a terminal.' Is this still realistic?" Realistically, someone needs to run the one-line install command and type the first prompt for them — a five-minute phone call covers it. After that, Codex's plain-English back-and-forth is closer to texting a very capable IT-savvy grandkid than to using a terminal in the traditional sense.

Related on explainx.ai

• Ethan Mollick let Codex loose on his X bookmarks for two hours — his other recent Codex story
• Codex $HOME deletion: GPT-5.6, full access, and the July investigation — the safety incident behind this post's guardrails
• 3 AI skills that matter most in late 2026 and 2027: Skills, loops, MCP
• AI skills every developer needs in 2026: the full roadmap
• Top 10 open- and closed-source agent harnesses in 2026
• OpenAI Codex plugin support: Claude Code guide
• Loop engineering for coding agents: Claude Code guide

Official: OpenAI Codex · Codex CLI documentation · Munder Difflin on GitHub

Codex CLI version numbers, flags, and sandbox behavior reflect the tool as of August 2026 and change frequently — check the official docs before relying on specific flag names. Munder Difflin details reflect its v0.4.4 release; it's an early-stage open-source prototype, not an official OpenAI product.