So How Do You Actually Vibe Code? Not the GPT App — Claude Code or Codex

Bottom line first: pasting code into a ChatGPT window? That’s not vibe coding — it’s hand-coding with AI seasoning. The real thing is an agent that reads the files on your machine, runs commands, and deploys. Two options do it: Claude Code or Codex. Run them locally as a desktop app or an OAuth CLI.

Side by side

	ChatGPT app/web	Claude Code	Codex
Reads/edits your files	❌ (copy-paste)	✅	✅
Runs terminal commands	❌	✅ (with approval)	✅ (with approval)
Deploys (clasp, wrangler)	❌	✅	✅
Login	account	Claude subscription OAuth / API key	ChatGPT subscription OAuth / API key
Form	web · app	CLI · desktop · web · IDE extension	CLI · desktop

The ❌ column is exactly what “copy-paste hell” is. An agent turns the whole column into ✅.

Why the GPT app falls short

ChatGPT’s web/app can’t see your folder. So the flow becomes:

get code → paste → save → run the command → error → copy the error → paste it back and ask → more code → paste again…

That round-trip is the opposite of vibe coding. An agent sits in the same folder, edits files directly, runs commands directly, reads the error directly, and fixes it. You only say “do this” and “yes / no.”

Option 1: Claude Code (Anthropic)

npm install -g @anthropic-ai/claude-code
cd ~/my-project        # the folder to work in (mkdir an empty one if needed)
claude                 # run

On first run, type /login and a browser opens for Claude-account OAuth. On a subscription you need no API key. The same thing exists as a desktop app (Mac/Windows), web at claude.ai/code, and VS Code / JetBrains extensions.

Handy bits:

• /init — have it scan the project once to build context (do this once at the start).
• An approval prompt appears before it runs a command or edits a file. Look, then y.
• claude mcp add ... — attach external tools (MCP) like gws-cli → Workspace access.

Option 2: Codex (OpenAI)

npm install -g @openai/codex
cd ~/my-project
codex                  # run, then sign in with your ChatGPT account

There’s a desktop app too. You pick how much it can do on its own with a mode — Suggest only / auto-edit but approve commands / Full-Auto. Start in an approval mode, then loosen it as trust builds.

What a first session actually looks like

Say you start in an empty folder. After running claude (or codex):

1. You: “Build an Apps Script web app in this folder. Log form submissions to a sheet and email me. Deploy it with clasp.”
2. Agent: shows you it’ll write Code.gs, index.html, .clasp.json → you approve.
3. Agent: asks to run clasp push → approve. (First time, it has you clasp login first.)
4. On an error, the agent reads and fixes it. (e.g. “Apps Script API not enabled” → it gives you the link to turn it on.)
5. You: “Add email-format validation” → one feature at a time, small and often.

The key move: when stuck, paste the error verbatim. You don’t debug — you let the agent read and fix it.

Which one?

• Already on Claude → Claude Code. Already on ChatGPT → Codex. (Reusing a subscription is the cheapest path.)
• New to both? Either. The workflow is nearly identical. Run one for a month and keep the one that fits.
• Want it inside an IDE (VS Code)? The Claude Code extension is convenient.

Watch out

In one line

Pasting is hand-coding. Log into Claude Code or Codex over OAuth and run it locally. That’s the real thing — files, commands, deploy, all done by the agent — and reusing a Claude/ChatGPT subscription means no keys and no extra cost.

Questions?

Stuck on which to start with? Just ask. Next step — getting what you built online → Stuck on localhost — now deploy it. The full picture → Vibe Coding 101.