Codex CLI vs Codex App: Which OpenAI Surface?

Codex CLI vs Codex app is the choice between OpenAI's terminal coding agent and the ChatGPT-hosted Codex UI. The CLI runs against your working tree, your shell, and (when configured) local MCP tools. The app runs in a browser or desktop shell, often with a cloud workspace, and it is the thing people mean when they say "I asked Codex in ChatGPT."

They are marketed as one product line. They do not share your unsaved buffer. Pick the surface that matches the job, then give it named screen context if the job is a UI bug. Codex CLI is the terminal deep dive. Claude Code vs Codex is the cross-vendor one.

What is Codex CLI, exactly?

Codex CLI is a command-line agent: you launch it in a repo, you prompt, it reads files, runs commands you allow, and writes diffs.

It feels like Claude Code. That is the honest comparison, not "ChatGPT with extra steps."

Strengths: your git state, your hooks, your local tests, no upload of the whole machine unless you configured that. MCP can be added so the CLI calls a local capture server on 127.0.0.1.

Limits: you have to like a terminal. Approvals for commands are on you. A bad rm policy is a bad day.

Install and auth follow OpenAI's current CLI docs; I will not paste a flag that will be wrong next month. Log in, cd to the repo, start a session.

What is the Codex app?

"Codex app" here means the ChatGPT Codex experience: a UI in ChatGPT (and any desktop wrapper OpenAI ships) where you assign a coding task, watch a cloud or connected workspace work, and pull a PR.

Strengths: no terminal anxiety, shareable threads, good for "go fix this repo I connected." Product people can open it.

Limits: the workspace is not always your laptop. Local unsaved files are invisible. MCP on 127.0.0.1 may not exist in that VM. A Mac Accessibility capture of your screen is a paste or a download unless you stay on a local CLI.

If you only ever work in the hosted UI, treat the Codex app as a remote intern. Brief it like one.

Codex CLI vs Codex app, side by side

Codex CLI

Runs: your terminal, your cwd Sees: local files, commands you approve MCP: yes, local servers Best for: daily implementation in a repo you already have open Weak at: non-developers who will not install a CLI

Codex app (ChatGPT Codex)

Runs: hosted UI, connected or cloud workspace Sees: what that workspace synced MCP: remote or none, depending on the week's product Best for: assigning a ticket from a browser Weak at: live UI on your laptop, unsaved edits, local-only secrets

Claude Code (for calibration)

Runs: terminal, local Sees: your repo MCP: first-class; see Claude Code MCP Best for: the same job as Codex CLI Weak at: people already paid for ChatGPT-only

No four-column spreadsheet. Those three cards are the decision.

How do I give either surface a UI bug that is not a riddle?

Pixels in the chat are the default and the trap.

On the CLI, register a local MCP server. PinVari → Connect → Codex when the app offers it, or point the CLI at $HOME/.pinvari/mcp/pinvari-mcp the same way you would any stdio server. The Mac app must be running; the connector uses 127.0.0.1:3402.

claude mcp add --scope user pinvari -- "$HOME/.pinvari/mcp/pinvari-mcp"

That command is the Claude-family installer. For Codex CLI, add the same binary in its MCP config. Do not invent a bare codex mcp add pinvari if the docs do not show it; the binary path is the durable part.

Then ⌥⌘A, circle, speak. The CLI pulls pinvari_next_instruction: role, label, frame, confidence, circled-versus-dwelled, per-mark words, crop, window text up to 40,000 characters, URL from AXWebArea.

On the Codex app, you usually cannot open that socket. Capture anyway. Paste the resolved text and the crop into the thread. You still beat "the button is weird."

pinvari_request_capture only helps a local agent. Hosted Codex cannot light your notch island unless a bridge exists. Do not pretend it can.

Which one should I use this week?

Use Codex CLI if you are a developer, the repo is on disk, and you want MCP or tests. This is the peer of Claude Code.

Use the Codex app if you are assigning work from a phone or a PM laptop without a checkout, and the repo is already connected.

Use both only with a rule. CLI owns in-progress branches. The app owns "start a PR from main while I am in a meeting." Otherwise you will merge two stories that never saw each other.

Use neither if the job is "name this control." That is a capture tool, then whichever agent you already pay for. One-time PinVari license versus another monthly seat is a separate budget line.

Failure modes I would budget for

The app edited an old SHA. The workspace had not pulled. You "fixed" yesterday.

The CLI ran a command you would not have run. Tighten approvals. Read the diff.

Both got the wrong button. You sent a full-page PNG. Send a name.

MCP shows zero tools. Bad path, app quit, or you configured the hosted app. Restart the Mac app. Use an absolute command path.

Electron preview with empty AX. Expect OCR fallback and low confidence. Ask, do not let Codex click-guess.

A practical split for a Mac team

Developers: Codex CLI or Claude Code, plus point-and-speak. Same MCP. Same Command Center.

PMs: Codex app for "write the first PR description," and PinVari filed to GitHub or Linear for the actual UI note. Do not make PMs install a CLI to report a broken checkbox.

QA: template plus capture. Assignee can be a human or a CLI session.

Intel and Apple Silicon both run the capture app. macOS 14+. Notarized DMG.

How do I move a task from the Codex app back to Codex CLI?

Download or pull the branch the app opened. Do not copy files out of the UI by hand if a PR exists.

Open that branch in a terminal. Start Codex CLI (or Claude Code) in that cwd. Replay the capture from the Command Center so the local agent sees the same named element.

If the app already wrote a wrong fix, revert that commit before you stack a local one. Two Codex surfaces both "helping" is how you get a button that is full width and also gone.

One branch owner at a time is the only merge rule that survives this comparison.

FAQ

Is Codex CLI the same model as the Codex app?

Often the same family, not the same runtime limits, tools, or context window policy. Judge them by the loop you get, not by the word Codex on the landing page.

Can the Codex app call a local MCP server on my Mac?

Assume no until you have proved a tunnel you trust. Local MCP is a CLI and desktop-IDE feature. The hosted app sees the workspace it was given.

Codex CLI vs Claude Code: should I pick for this comparison?

If the question is CLI vs app inside OpenAI, stay there. If the question is which terminal agent, read the Claude Code comparison and pick the model you already have keys for.

Which is cheaper, Codex CLI or the Codex app?

Both meter OpenAI usage on your plan. The CLI can waste less if you send structured captures instead of huge screenshots. The app can waste more in a chatty thread. Check your account; I will not invent a price.

Can I use voice with Codex CLI?

Yes. Dictate into the prompt, or use ⌥⌘V / ⌥⌘A and let MCP deliver the transcript bound to an element. The CLI does not ship a Mac overlay of its own.

Should a founder use the Codex app instead of the CLI?

If you will not open a terminal, yes. File UI bugs as named captures so the app does not restyle the wrong card. Graduate to a developer and a CLI when the repo needs tests.