说个我自己都没料到的事,Codex现在能完全免费用了。
而且是个能长期用的法子,文本、图像、视频全都包。方法我前前后后折腾过几遍,能稳定接上,今天就把它拆开讲给你。
核心思路特别简单,把Codex背后的模型,换成一个能力在线、又长期免费的大模型,Agnes。
这名字你可能没听过,我之前也完全没印象。
后来扒了一下才知道,是家挺低调的新加坡公司。官方说它是全球TOP10的AI实验室,文本、图像、视频三块都排在前面。
我花了点时间琢磨,成功把它接进Codex了。
之前不少人遇到的,接别的模型后插件全没了的问题,这次也能绕开。接完之后,Codex的插件、Skill、聊天记录,全都还在。
整套流程只用到一个工具,Codex++。中间有几个容易踩的坑,我边讲边提。
第一步,先把Codex装好
已经装过的可以跳过这步,只要把Codex彻底退出就行。
还没装的话,直接搜索Codex官网,挑适配你电脑的版本下载。
用免费API登录还有个附带的好处,门槛降下来了。
以前登Codex得手机号验证,或者充个Pro才进得去,现在连OpenAI账号都可以不用。
装好以后先打开Codex,会蹦出登录界面,别管它,直接关掉,然后把Codex完全退干净。
第二步,装Codex++
下载地址在这
https://github.com/BigPizzaV3/CodexPlusPlus/releases
Mac用户如果打开报错,终端里输入这两行
sudo xattr -rd com.apple.quarantine "/Applications/Codex++.app"sudo xattr -rd com.apple.quarantine "/Applications/Codex++ 管理工具.app"
这一步是在解除macOS的安全隔离,对没签名的第三方应用是常规操作,知道自己在干嘛就行
下完你会看到两个东西,Codex++和Codex++管理工具,两个都需要装上,后面会讲分别怎么用。
第三步,拿到Agnes的API Key
打开官网
https://agnes-ai.com/
先注册个账号,进控制台,找到API Key那一栏。新建一个Key,复制好存下来,待会儿要填。
最后一步,把Agnes接进来
打开Codex++管理工具,点左边的供应商配置,选添加供应商。
供应商名称,写Agnes。接入模型,选纯API。模型配置,填agnes-2.0-flash。
Base URL填这行
https://apihub.agnes-ai.com/v1
Key填刚才复制那串,模式选Chat。
然后点从上游获取,正常的话会拉到五个模型。填完记得保存。
接下来这步是关键。
配完之后,选中这个模型,点右上角的重启Codex++。
这个重启Codex++,等于帮你双击了一下Codex++这个工具本身,省得你再切回去手动双击。
再梳理一下流程。
换模型之前,先把Codex彻底退掉。
打开Codex++管理工具,选好模型配置。
点右上角重启Codex++,或者直接双击Codex++。
一定要注意,这里是双击Codex++,别自己手动去开Codex,不然接不进去。
因为手动开的是官方那个Codex,走的是官方登录态,自然读不到你刚配的API。得让Codex++来拉起Codex,配置才会注入进去。
看看效果如何
打开Codex,能看到用API成功登上了。
先随手测一句,问它今天几号。答对了,说明模型这边通了。
再让它在桌面建个文件夹,也建出来了,说明调本地工具这条路也通。
再检查一下插件,搜个hyper frames,该有的都在,跟用GPT登录时一模一样。之前接别的模型把插件弄没了的,也能在Codex里重新填一遍API Key找回来。
之前的Skill也能照常调用。下次再切回这个模型,原来的聊天记录也都给你留着。
右下角你会看到有图像、视频模型,但这俩在这儿点不动。
原因是Agnes确实有生图生视频的能力,可它对外只给OpenAI Chat Completions兼容的接口,图像和视频走的是另一套,靠切模型或者直接对话都唤不出来,得用Skill去调API。
简单点的玩法,是让另一个本地跑着的Agent帮你把Skill装好、把API Key写进去。我就让Claude Code帮我装了个生图的Skill,简单试了试,出图挺快,效果也还行。
再说下限制。
Agnes免费版现在主要卡的是速率,每分钟最多调20次,应付日常基本够用了。
官方在推1M的超长上下文,但1M上下文和4K图片这两块不会一次性全放开,是分批灰度的,接下来几天会陆续覆盖到所有人。
最后想说的话
Agnes 官方文档里有句话,我看到的时候很受触动:
要让世界级的人工智能属于每个人。
世界级的智能,不该只握在少数大机构手里,而应该成为每一个开发者使用、每一个产品实现集成、每一个用户实现受益的基础能力。
至于免费能撑多久,谁也说不准。
但有件事我越来越确定。
工具越是强大,普通人就越需要找到自己不可被替代的东西。
这点东西,就是给自己挖的护城河。
我是路人甲,前数据分析、产品人,现创业者。关注 AIGC 人工智能,分享实用的 AI 应用。让 AI 变成你触手可及的生产力。
Here's something I didn't expect myself: Codex can now be used completely for free.
And it's a method you can rely on long-term, covering text, images, and video. I've tested it back and forth several times, and it connects stably. Today I'll break it down for you step by step.
The core idea is very simple: replace the model behind Codex with Agnes, a powerful large model that's online and free long-term.
You might not have heard of it; I hadn't either.
After digging a bit, I found out it's a rather low-key Singapore-based company. Officially, it claims to be a top-10 AI lab globally, ranking at the forefront in text, image, and video generation.
I spent some time figuring it out and successfully integrated it into Codex.
This time, I also bypassed the common issue where plugins disappear after switching to other models. After integration, Codex's plugins, Skills, and chat history all remain intact.
The entire process only requires one tool: Codex++. There are a few common pitfalls along the way, which I'll point out as we go.
Step 1: Install Codex first
If you've already installed it, you can skip this step; just make sure to fully quit Codex.
If you haven't installed it yet, go directly to the official Codex website and download the version compatible with your computer.
Logging in with a free API also comes with an added benefit: the barrier to entry is lowered.
Previously, logging into Codex required phone verification or purchasing Pro. Now, you don't even need an OpenAI account.
After installation, open Codex first. The login screen will pop up; ignore it, close it directly, and then fully quit Codex.
Step 2: Install Codex++
Download link here
https://github.com/BigPizzaV3/CodexPlusPlus/releases
Mac users: if you encounter an error when opening it, enter these two lines in the Terminal
This step bypasses macOS's security isolation, which is a standard procedure for unsigned third-party apps. Just make sure you know what you're doing.
After downloading, you'll see two items: Codex++ and the Codex++ Manager Tool. Both need to be installed; I'll explain how to use each later.
Step 3: Get the Agnes API Key
Open the official website
https://agnes-ai.com/
Register an account first, go to the console, and find the API Key section. Create a new Key, copy it, and save it for later use.
Final step: Integrate Agnes
Open the Codex++ Manager Tool, click "Provider Configuration" on the left, and select "Add Provider."
Provider name: Agnes. Integration model: Pure API. Model configuration: agnes-2.0-flash.
Base URL: enter this
https://apihub.agnes-ai.com/v1
Key: paste the one you just copied. Mode: Chat.
Then click "Fetch from upstream." Normally, it will pull five models. Remember to save after filling everything in.
The next step is the key part.
After configuring, select this model and click "Restart Codex++" in the top right corner.
This "Restart Codex++" is equivalent to double-clicking the Codex++ app itself, saving you from having to switch windows and do it manually.
Let's recap the workflow.
Before switching the model, fully quit Codex first.
Open the Codex++ Manager Tool and select the model configuration.
Click "Restart Codex++" in the top right corner, or simply double-click Codex++.
Important: double-click Codex++ here. Do not manually open Codex yourself, or it won't connect properly.
Because manually opening it launches the official Codex with the official login session, which naturally won't detect the API you just configured. Codex++ must launch Codex for the configuration to be injected.
Let's see how it works.
Open Codex, and you'll see it has successfully logged in via API.
First, test it with a quick question: ask what today's date is. If it answers correctly, the model connection is working.
Then ask it to create a folder on the desktop. If it does, the local tool integration is also working.
Check the plugins again. Search for "hyper frames," and everything is there, exactly like when logging in with GPT. If you previously lost plugins by switching models, you can recover them in Codex by re-entering the API key.
Your previous Skills can also be called as usual. Next time you switch back to this model, your original chat history remains intact.
In the bottom right corner, you'll see image and video models listed, but you can't click them here.
The reason is that Agnes does have image and video generation capabilities, but it only exposes an OpenAI Chat Completions-compatible interface externally. Images and video use a separate pipeline and can't be triggered by switching models or direct chat; you need to use a Skill to call the API.
A simpler approach is to have another locally running Agent install the Skill and input the API key for you. I had Claude Code install an image-generation Skill for me, gave it a quick try, and the output was fast with decent results.
Let's talk about the limitations.
The Agnes free tier is currently mainly rate-limited, capping at 20 requests per minute, which is basically sufficient for daily use.
The official team is rolling out 1M context length support, but the 1M context and 4K image features won't be released all at once. They're being rolled out in batches gradually over the next few days.
Final thoughts
There's a sentence in the official Agnes documentation that really resonated with me:
"World-class artificial intelligence should belong to everyone."
World-class intelligence shouldn't be monopolized by a few large institutions. Instead, it should become a foundational capability that every developer can use, every product can integrate, and every user can benefit from.
As for how long the free tier will last, nobody can say for sure.
But one thing I'm becoming increasingly certain about is...
The more powerful tools become, the more ordinary people need to find what makes them irreplaceable.
That is your own moat.
I'm Passerby A, formerly in data analytics and product management, now an entrepreneur. I focus on AIGC and AI, sharing practical AI applications. Let AI become productivity at your fingertips.










