Codex is four different things — which one do you want?
“Codex” now names a feature inside ChatGPT, an open-source command-line program, a desktop app, and an IDE extension. People argue past each other constantly because of it — one person means a tab in a browser, the other means a binary on their machine. This page separates them, and says plainly which one you can run on hardware you control.
The short answer
The open-source CLI has not gone anywhere. It is Apache-2.0, sits at roughly 75,600 GitHub stars, and has shipped over 700 releases. Codex showing up inside ChatGPT was an addition, not a replacement — which is worth stating clearly, because the two get confused with a genuine deprecation that happened elsewhere this year.
The four surfaces
| Surface | Runs where | Open source | Best for |
|---|---|---|---|
| Codex in ChatGPT | OpenAI’s servers | No | Delegating a task from a browser, nothing to install |
| Codex CLI | Your machine or your server | Yes — Apache-2.0 | Your own files, scripting, CI, running on a VPS |
| Desktop app | Your machine | No | Local work without living in a terminal |
| IDE extension | Your editor | No | Staying inside the editor you already use |
The models behind them are the same. What differs is where the work happens, what it can reach, and whether you can automate it.
Why the CLI is the one that matters here
Only the CLI is a program you own a copy of. That has three consequences the other three surfaces cannot offer:
- It runs against your filesystem, on a machine you choose.
- It can be scripted, cron’d and put in CI, because it is just a command.
- It can live on a server, so work continues when your laptop sleeps.
Worth being honest about the limit: the CLI is open source, but it still talks to OpenAI’s models and authenticates with a ChatGPT plan or an API key. Open client, hosted brain. If provider independence is what you are after, a model-agnostic agent is the better fit — see the Gemini CLI retirement for why that distinction is not academic.
What the CLI costs to run on a server
We install Codex as a one-click preset and measured it on a 1 vCPU / 1 GB server:
- Install: under a minute — Node 22 plus one global npm package.
- Memory at idle: nothing resident. The box sits at its OS baseline between sessions.
- Minimum server: 1 vCPU / 1 GB, proven rather than estimated.
“Nothing resident” is the useful part: Codex is invoked per task rather than run as a daemon, so it costs you nothing while idle and can share a box with whatever else you are running. One caveat we will not hide — our measurement covers a box with Codex installed and ready, not one mid-task, because our automated checks install and verify it without driving a session.
Method and the rest of the numbers: VPS hosting for AI agents.
Picking one
- Just want it to do a thing? ChatGPT. Nothing to install.
- Working in your own repo? The CLI.
- Automating, or want it running while you are away? The CLI, on a server.
- Want to not depend on one vendor? None of them — look at a model-agnostic agent.
Related
- Run Claude Code & Codex on a VPS — the step-by-step version.
- Gemini CLI was retired — what happens when the vendor half of an open client goes away.
- VPS hosting for AI agents — measured comparison across every agent here.