The Ultimate Guide to Using Codex

A comprehensive guide for all beginners who want to seriously use Codex.

Just the other day, my high school buddy who runs a car insurance business came asking me how to use Codex. Looks like this product is spreading fast beyond its initial audience.

Honestly, I still strongly recommend everyone try Codex. What OpenAI makes is really good and powerful.

Beyond writing code, Codex can do a lot more than most people imagine. Things like generating PPTs, data analysis, exam preparation, verifying business data — it handles them all.

Let’s skip the fluff and get started.

What is Codex?

Codex is an Agent product launched by OpenAI with a great user experience. To put it simply, many similar domestic products are iterating by following Codex’s lead.

But for beginners, there’s one hurdle to using Codex: you need a ChatGPT membership.

The membership is expensive — $20 per month. Plus, you can’t buy it directly because it doesn’t support domestic payment methods.

I previously recommended connecting DeepSeek V4 to Codex since DeepSeek is very cost-effective.

However, it doesn’t support multimodality. If you throw in an image, DeepSeek V4 can’t recognize it at all — it throws an error, like in the screenshot below.

Also, V4 is weak in Office-related tasks — for example, making a PPT that looks decent is hard.

So recently I switched to Kimi’s model. Let me put it this way: I think Kimi is probably the best model for connecting to Codex domestically.

It has several advantages: first, it natively supports multimodality, accurately understanding images and videos.

After all, this kind of need is super common in daily life — it’s normal to screenshot something at work to communicate.

And for non-coding users, a very important daily scenario is handling Office tasks or building web pages.

Kimi excels in generating PPTs, spreadsheets, and web pages — everything looks really polished with good aesthetics. For example, the table below was made by Kimi — it even froze the first two header rows.

How to connect Kimi in Codex?

So how do you connect Kimi’s model in Codex?

It’s simple. First, download the Codex client — just go to the OpenAI official site. There’s a macOS version and a Windows version.

https://openai.com/codex/

After downloading Codex, the overall steps are:

1. Download the configuration tool CC Switch, which helps Codex connect to Kimi’s model.
2. Go to Kimi’s website to buy a membership (very cheap), and configure it in CC Switch.

Codex is the software you ultimately want to use. Kimi is the model you want Codex to call. CC Switch is the bridge that handles the connection and configuration.

Once you understand what each of these three things does, the installation process is basically just clicking “Next”.

I remember someone asked if you need to log into Codex after downloading. No, you don’t. After downloading Codex, follow my steps and continue installing CC Switch.

This involves no Codex login, no messing with phone numbers.

One more thing: some users who have used Codex might be confused — hasn’t Codex already supported third-party model connections? Why do we still need a bridging tool like CC Switch?

Because Codex requires third-party models to support its own Response protocol, but currently, Kimi doesn’t support it.

I asked a friend at Kimi, and they said it should be coming soon. Once it’s supported, CC Switch will no longer be needed.

After installing Codex, we proceed to install CC Switch. Here’s its official site:

https://ccswitch.io/

Click the “Free Download” button — it will redirect to a GitHub page. Don’t bother with the complex content; scroll all the way down.

At the bottom, you’ll see many installation packages. For Windows, choose the one ending with .msi. For Mac, choose the .dmg package.

After installing CC Switch, you’ll see an interface like this:

On the left, be sure to select the icon with the OpenAI logo (which I’ve highlighted in red). The other icons are for configuring other software.

Click the “+” button on the right. Once inside, you’ll see a list of model providers.

No need to study them all — we’re connecting to Kimi, so choose “Kimi For Coding”. Note it’s “Kimi For Coding”, not just “Kimi”.

Next, you’ll start configuring the API Key.

What is an API Key? Let me explain with an analogy, and also clarify what people often call tokens.

Imagine the large model as a small power plant, and tokens are the electricity it generates. Every time we ask AI to do something, we’re essentially using electricity.

OpenAI, Anthropic models are great too, but the electricity bill is relatively expensive.

Kimi’s advantage is its high cost-effectiveness, making it suitable for ordinary users to handle everyday office tasks.

An API Key can be understood as our electricity card number. After buying a Kimi membership, you’ve essentially purchased some electricity.

Then you fill the API Key into the configuration tool, and Codex knows that when it calls Kimi, it uses the electricity from this card.

So never leak your API Key.

You can buy Kimi membership at the link below. The cheapest tier is around 40 yuan per month — if you’re just starting out, you can buy this one without too much financial pressure.

Once you use it more and confirm the model works well and helps you solve work problems, you can upgrade to other tiers.

https://www.kimi.com/membership/pricing

After getting the membership, go to the console page and create a new API Key. This API Key is the electricity card number we mentioned.

Later, you need to fill this API Key into CC Switch so Codex knows to deduct from your account balance.

After configuring, click “Add” — Kimi’s model is now connected to CC Switch.

Next, we need to click the settings button in the upper left corner of the main page to configure routing so Codex can recognize it. Refer to the image below.

Enter the settings page, select the routing menu, then turn on all three switches highlighted in red in the image.

Finally, return to the main page and enable “Kimi For Coding”.

And that’s it — done. Open Codex again, send it any message, and you’ll be pleasantly surprised that it can reply. That’s the magic of Codex.

Specific Use Cases

Codex is not hard to pick up — just play around with it and you’ll learn.

Let me give an example. Around mid-year, many people probably need to make PPTs. We can give Codex the relevant files and data and ask it to create a PPT according to our requirements.

Like below, some are screenshots, some are PDFs — it doesn’t matter; it understands all of them, since we connected Codex to Kimi’s multimodal model.

Here’s the PPT I got just now — isn’t it nice? Of course, the AI’s first output won’t be 100% perfect. Then we can directly modify it. If the changes are big, we can continue to give Codex new instructions.

In short, after installing Codex, treat it as a colleague who can do work.

I strongly recommend putting a sticky note by your computer: for everything, first think whether Codex can do it.

A few more examples.

At work, I often need to convert PDF to Markdown files — I can give that to Codex.

Also, the other day I had a family member’s medical checkup report in PDF. I just threw it in and asked Codex to analyze the report and compile a summary of recommendations into a Lark document — all done.

Another example: everyone runs out of computer storage, right? You install some weird software, and when you try to uninstall it, it doesn’t clean up completely.

You can give that to Codex too. Say, “Help me check if a certain software has been completely uninstalled.” It will find files and dependencies related to that software and see if any leftovers remain.

If there are any, just ask it to clean up.

A while ago, I freed up 20GB of space on macOS using this method.

I asked Codex to check if all recently installed apps were completely uninstalled. It did the analysis, listed them out, and I asked it to handle them as needed.

Almost everything at work, you can try giving to Codex first.

Besides telling Codex what to do and letting it execute, often we have our own workflows that can be standardized.

For example, my company stores all data in Lark Base. I installed Lark’s CLI in Codex so it can directly interact with Lark.

Every morning, I need to check yesterday’s payment collection data and see if the numbers match across several tables. This used to be manual and easy to forget. Now I created a Skill.

What is a Skill? It’s basically standardizing a set of work methods or experiences into a file so the AI knows how to do it.

For example, with account reconciliation — what accounts to reconcile and how — that was only in my head. I told Codex to create a Skill for me, and now it knows how to reconcile.

After creating the Skill, I set up a scheduled task. Codex has a scheduling feature, so it can run automatically every day to check for anomalies in payment collection data.

Another example: I recently bought a desktop e-ink device. I used Codex to connect it with my Lark Calendar.

Codex pushes my Lark Calendar to the e-ink device at fixed times in the format specified in the Skill.

So I can just glance at my desktop to know upcoming tasks and to-dos — no need to check my phone or computer calendar anymore. One less context switch.

Skills are not difficult — don’t be intimidated by new terms.

Let me tell a story to help you understand what a Skill is.

Years ago, I was chatting with a respected older woman who said: many people are sloppy, fixated on details, but never focus on optimizing the process.

Process has a huge impact on outcomes. Doing A then B vs B then A — there’s a big difference.

She gave a cooking example: after stir-frying, she immediately wipes down the stove. Because the oil splattered from the pan hasn’t solidified yet — cleaning while hot is much easier than cleaning after the meal.

Another example: during cooking, in any spare moment, she washes utensils that won’t be used later, reducing the post-meal cleanup burden. That’s process.

I was deeply moved. Everyone has a different workflow and method for doing things. These directly affect the final quality and efficiency.

A Skill is essentially encapsulating that method into a fixed file so the AI knows the specific process.

How to create a Skill? No technical details needed. Just tell Codex, “Create a Skill for data analysis,” and then describe your workflow in natural language.

For this kind of scenario, I highly recommend using voice input. Imagine you’re talking to a colleague and just speak the workflow naturally.

You don’t need to be perfectly precise — the AI can understand. If something is unclear, interact a few more rounds, and eventually you’ll have a Skill.

Next time you do the same work or set a scheduled task, just ask Codex to run that Skill.

To summarize:

First, let Codex handle things first. Develop this habit, and you’ll gradually find that many things at work can be delegated to AI.

Second, deliberately standardize repetitive workflows into Skills.

Third, if a task needs to be done daily or weekly, set it as a scheduled task.