It’s been six months since our last beginner’s guide to Codex. In those six months, Codex has evolved from a standalone desktop app into a feature integrated inside the ChatGPT client, and filled out a whole set of new capabilities: Sites, the office productivity suite, App Shots, Hooks, and more. This post breaks down every key update and focuses on the most practical “token-saving toolkit” for anyone currently using — or about to use — Codex.

Bottom line up front: if you only use Chat mode to talk with Codex, the biggest win in this version is essentially “getting the ChatGPT desktop client for free.” If you use Codex mode to run engineering tasks, the newly added Hooks, remote control, and multi-folder projects turn it into a real production-grade Agent. Let’s go through them one by one.

1. Desktop Integration: How to Choose Between the Three Modes

The new ChatGPT desktop client has folded Codex in. You can switch between three modes in the top-left: Chat / Worker / Codex. The differences between them can be summarized in one table:

Mode Positioning Can it touch the local machine by default? Does it consume subscription quota?
Chat Conversational Q&A Agent (explain code, translate, look things up) No None
Worker General-purpose office Agent (research, docs, spreadsheets, PPT, reports) Yes Consumes
Codex Coding and engineering Agent (edit code, run tests, fix bugs) Yes Consumes

On the UI side, Codex mode adds three things over Worker: Git branch management, a built-in terminal, and a file tree. The sidebar only shows local work conversations, with denser code-related features. Worker mode hides Git-related entries entirely, making it more suited to daily documentation work. Both engineering modes draw down the 5-hour quota and the weekly quota — once either is exhausted, you can no longer use the feature. That is the fundamental pain point every token-saving tip below aims to solve.

2. 4 Hard Tips That Actually Save Tokens and Quota

Tip 1: Talk it through in Chat mode first — do all the planning at zero cost

Chat mode does not consume any Codex subscription quota. That’s a free card you’re dealt. The right way to play it: when a task comes in, talk through the requirements in Chat first, then switch to Codex mode to execute. A typical flow:

  1. In Chat mode, ask Codex to design an HTML video editing console and crank the reasoning effort to the highest setting.
  2. Have it search for recent tech frameworks and output a Markdown design document.
  3. Switch to Codex mode, create a new project, @-reference the previous conversation (the new version supports cross-project @), and type “execute this plan.”

In testing, this approach saved at least half the tokens. The logic is simple: upfront research and requirements alignment don’t need to burn the expensive execution quota. Handing it off to free Chat costs you nothing.

Going further, if you already have a project on your machine, you can use a plugin to let Chat mode read and write local files directly, achieving local file operations without spending a single quota unit:

  1. Open the ChatGPT web app in your browser → Settings → Plugins → search for and install Remote Desktop Commander.
  2. Copy the one-line install command provided by the plugin and run it in your local terminal to complete pairing.
  3. After that, in Chat you can ask GPT to read directories, read files, and review code — just like in Codex.

There are also open-source projects on GitHub like codex-as-position-to-chatgpt and codex-chatgpt-web that do similar things. The principle is the same: route planning/review tasks to the web ChatGPT so Codex only handles the actual execution.

Tip 2: Offload the dirty work to the Luna small model

The GPT-5/6 Luna model has been priced down again compared to the previous generation — it’s now cheaper than DeepSeek-V3.1 Flash (for how model choice affects cost, see this deep dive on how large models work). The judgment is straightforward: using Luna is roughly equivalent to unlimited usage. So simple work — scheduled tasks, Skills that need to run many times, browser automation — can all go straight to Luna.

Switching is easy: pick Luna from the model selector and set the reasoning effort to “High” or “Very High”. Only genuine programming work or genuinely hard creative tasks need GPT-6 Soul or Astra. When Luna doesn’t deliver, refining your prompts is usually more effective than switching models.

Tip 3: Plan mode + GrooveMe — do the planning upfront

Anyone who uses Codex regularly would probably agree: “If you could only pick one capability out of all of them, it would be Plan mode.” The real value of Plan mode isn’t “auto-writing code” — it’s thinking the solution through so you don’t have to redo work after you’ve started. The recommended flow is:

  1. Turn on Plan mode, pick a powerful model, and crank the reasoning effort to maximum.
  2. Let Codex use the follow-up window to actively align on requirements with you (it typically asks three key questions).
  3. Once the full plan is out, hit “execute this plan.”

To make Plan even more thorough, install the GrooveMe Skill (rebranded from its previous name). It keeps asking follow-up questions until the entire solution is thought through. To install GrooveMe:

1
2
3
4
5
# macOS
~/.codex/skills/

# Windows
C:\Users\<username>\.codex\skills\

Copy the entire downloaded GrooveMe folder into the skills/ directory, restart Codex, and type / to invoke it. Combined with Plan mode, in our testing Codex asked the user a total of 22 questions before touching any code, working out every detail of the project first. Once you have a detailed plan, during actual coding you can actually lower the reasoning effort from “Maximum” to “High” or “Medium” — second-guessing decided code is pure token waste.

Tip 4: Activate at dawn to get an extra quota refill during the workday

The 5-hour window starts counting from the moment you first open Codex. If you open at 9 AM, the reset times are 2 PM and 7 PM, so you only get two refills in a workday. But if you open it at 5 AM, the reset times shift to 10 AM and 3 PM — giving you three refills during normal working hours.

You don’t actually have to wake up early for this. Create a local scheduled task in Codex, set the execution time to 5 AM every day, and keep the task content as simple as possible (for example, have Luna say “hello”). One key point: the execution location must be set to “This device” — cloud-based tasks don’t consume subscription quota, so they won’t trigger the activation.

3. Route Your Subscription Quota Into Your Own Project (Pi AI)

If you already have an official Codex/ChatGPT subscription, you don’t need to set up a separate API key — you can drive your own tools directly with your subscription quota. Pi Agent has open-sourced the @pi-ai Node package:

  1. Visit Pi’s GitHub page → Packages → the AI package, which wraps dozens of model providers on the market into a unified calling spec.
  2. Feed the AI package’s README link to Codex and tell it “I want to build a ‘Pictionary’ web game — a canvas for users to draw on, and the AI guesses what it is.”
  3. Have Codex follow the Pi AI docs and add a “Model Subscription” configuration option in the page.
  4. Once the project is done, click “Connect Model” in the top-right of the page → select “OpenAI Codex Subscription” → pick a vision model like GPT-6 Soul → click “Connect Subscription Account” and authorize via browser login.

The end result in our test: the user draws a few strokes on the canvas, and the AI guesses it directly. The entire flow uses zero API key balance — it’s all subscription quota. One subscription covers all your AI needs.

4. Use CC-Switch to Plug in Domestic Models

If you don’t have a Codex subscription or you’ve run out of quota, you can plug open-source domestic models into Codex. The core tool is CC-Switch:

  1. Go to the CC-Switch website, choose Codex at the top, and click + to create a new config.
  2. Pick DeepSeek (it has dedicated optimizations for the OpenAI Response API and performs best).
  3. Go to the DeepSeek website to create an API key, paste it back into CC-Switch, and click “Add”.
  4. Quit Codex, enable DeepSeek in CC-Switch, then restart Codex.
  5. After that, you’ll see DeepSeek’s two models available in the model selector.

When you need to switch back to the official subscription, just switch back to “Official Subscription” in CC-Switch, enable it, and restart.

5. Sites: Deploy Shareable External Apps at Zero Cost

Codex’s Sites feature lets you turn your requirements directly into a shareable external app with a public domain, complete with frontend, backend, database, object storage, domain, and environment variables. A concrete example is a “Hiking Check-in for Office Workers” app:

  1. In the conversation window, @-select size (the Sites plugin), and type the requirement: upload photos to object storage, extract coordinates from EXIF to locate the peak, drop pins on a map, store all check-in records in a database.
  2. Codex first writes the code locally, then uses the plugin to deploy the code to the cloud, delivering a public link directly.
  3. Bottom-left sidebar → “Sites” → find the published app → “Settings” lets you see database records, visitor analytics, and custom domain.

Custom domains go through Cloudflare and require three records:

Type Name Purpose
CNAME Your subdomain Points to the target assigned by Codex
TXT _acme-challenge subdomain Certificate verification
TXT Subdomain Domain ownership verification

Once added, tick “I’ve added the records at my provider” in Codex settings and the binding is done. Sharing permissions can be set to “Anyone on the internet can use it” — anyone with the link can access it. Sites is one of the most worthwhile features in this update: a single line of conversation gets you a real app with a database.

6. The Four Office Productivity Tools: Word/Excel/PPT/PDF

Codex comes with four built-in office plugins: @documents (Word), @spreadsheet (Excel), @presentations (PPT), plus a PDF tool. A typical scenario:

A teacher has two documents on hand — this semester’s class schedule (with classes taken but lacking details marked in red) and a lesson plan template. She wants Codex to fill in activity plans for an entire semester based on the template.

In practice, Codex called the Document plugin to read the red text, wrote a Python script to batch-generate Word files, and delivered a complete activity lesson plan document directly. If you’re not happy with the result, select a paragraph and click “Request Changes” to have Codex revise it as needed. Or pick “DefaultApp” from “Open With” to keep editing in your local Word.

The Document plugin also has a Create Template feature: upload a Word file as a template, and future invocations of the plugin can create files based on that template’s format. Perfect for scenarios with fixed document layouts.

The PPT plugin’s templates look stiff and aren’t pretty enough. An alternative: PPT Master on GitHub, developed by “和雨果”, generates all images automatically using Codex’s image generation capability. It looks noticeably better than the built-in templates — a good fit if you care about how your slides look.

7. App Shots: Let Codex See What You’re Working On

App Shots are triggered by default with two Alt keys (or two Command keys on Mac). When you send Codex an App Shot, it doesn’t just get a screenshot — it also submits every UI control on the screen as structured text. The combination of text and image lets Codex understand the software’s state much more efficiently.

A concrete scenario: “I need to add a black border to the selected text in PowerPoint but don’t know how.” Press the shortcut, send the snapshot to Codex, and Codex gives you step-by-step instructions directly. This kind of “see it once and understand” interaction is far more efficient than spending ages describing “what I currently have open.”

8. September 26 UI Overhaul: Sidebar Restructured

The September 26 update split the sidebar into three sections:

  1. Leftmost function menu: scheduled tasks, library, images, plugins, Skills, and more.
  2. Middle conversation list: centrally manages all chats.
  3. Right main interface.

A few features worth pulling out:

  • Images: the image generation feature has been spun off as its own entry. You can generate images from text alone, or sketch first and have Codex render from the sketch.
  • Customize: a centralized view of installed plugins and Skills. You can temporarily disable unneeded Skills to save context, or uninstall them entirely.
  • Library: a centralized view of all files uploaded or generated in Chat mode — images, text, and research reports are all searchable. Images support text-within-image search (in our test, searching for “office four-pack” got hits).
  • Maps: the intent was a life-style entry point (restaurant recommendations, route planning), but the current bugs are severe — every query errors out directly, and the data is mostly US-based. It can barely be used on the ChatGPT web; the desktop client is best avoided for now.

9. Hooks: A Safety Net for Dangerous Operations

Hooks is an extension mechanism Codex introduced. While the Agent is running it exposes a series of events (pre-tool-use, permission-request, post-tool-use, etc.), and you can attach your own scripts or MCP tools to these events. A real-world case:

Someone reported on the forum that Codex mistakenly wiped out their entire C: drive, so now they need a Hook to prevent bulk deletion.

Just have Codex write a Hook:

  • In the pre-tool-use event, check whether the current command is a bulk delete (rm -rf, Remove-Item -Recurse, etc.).
  • If so, block the command from running and instruct the AI that it must delete files one by one or stop and let the user handle it manually.

Once Codex finishes writing the Hook, find it in Codex Settings → Hooks and click “Trust” to activate it. In our test, when Codex was asked to delete the contents of a folder, the Hook successfully blocked it and only allowed single-file deletion. Hooks aren’t as flexible as Pi Agent’s plugins, but they’re much simpler to configure — perfect for this kind of safety guardrail.

10. Goals Mode / Computer Use / Recorder: Three Underwhelming Features

The remaining three features are all underwhelming. Quick summary:

  • Goals mode: theoretically suited for scenarios where “the task is big enough, the goal is clear enough, and multiple rounds of iteration are needed” — but current models are already strong enough to push forward autonomously as long as your prompts are clear. In our testing, it was hard to find a single scenario that absolutely required Goals mode.
  • Computer Use: lets Codex operate any software on your computer, but on Windows it seizes the keyboard and mouse, is painfully slow, and burns tokens at a ridiculous rate (10% of the 5-hour quota in a few minutes). The recommendation: use MCP or CLI instead of Computer Use whenever possible.
  • Recorder and Replay (Mac only): records your actions as a Skill, but the actual execution still falls back on Computer Use, so it burns tokens the same way and has no Windows version. Practical value is very low.

11. Remote Control and Project Partitions: Two Hidden Gems

Two final, very practical features deserve a callout:

  • Remote control: sidebar → “Remote Control” → “I’m signed in on desktop” → bring up the pairing code on the desktop Codex, scan it to bind, and your phone can then control Codex on your computer. Going further, you can authorize a Windows Codex to control another Mac in the Windows Codex settings, then pick the remote device when creating a project and the task will run on the remote device — cross-device task dispatching at its most useful.
  • Project partitions: Codex’s sidebar groups conversations by project, and now you can add another layer above projects called “partitions”. For example, create a “Self-media Tools” partition and group related projects inside it for cleaner management.

Final Thoughts

If you could only take three things away from this update, they’d be these three:

  1. Plan at zero cost in Chat mode first, then switch to Codex mode to execute — you can save more than half your subscription quota.
  2. Offload all repetitive, low-difficulty tasks to Luna — it’s roughly equivalent to unlimited usage.
  3. Plan mode + the GrooveMe Skill are the standard kit for complex engineering projects. Think the solution through before you write code — saves tokens and reduces rework.

Overall, the foundation of this version of Codex is already solid — Sites, the four office plugins, App Shots, and Hooks are all four solid and useful. But the decorative, exploratory features are still excessive (Goals mode, desktop pets, Computer Use, Recorder) — just pick the useful ones when you use it; no need to turn them all on.

The hands-on details for site deployment, CC-Switch configuration, and routing your subscription quota into Pi AI only got the core steps here — actually running through it will probably take another 10-20 minutes, and you’re welcome to share any pitfalls you hit in the comments.

If you’re more interested in the big picture of the recent AI industry, check out this 10 Big AI Industry Stories of the Week; for daily news, the AI Daily series updates every day.