What ChatGPT can't do, Codex App can; what ChatGPT can do, Codex App does better.
As one of the most powerful AI Agent products currently available, Codex App can help you achieve everything you want: code development, reading and writing files, deep research, and more.
After reading this article, you will understand all the basic concepts of Codex App, including:
- What is Codex App? Why is it so powerful?
- How to download and install Codex
- A plain-English explanation of Codex's core functions
- Essential Codex settings
- Recommended practical tips for Codex
To help readers read more effectively, I will update with advanced gameplay later, taking you step-by-step to unlock the mysteries of Codex and stimulate your curiosity about AI.
I. What is Codex App? How is it different from ChatGPT?
Codex App is an AI Agent desktop application launched by OpenAI. If past AI was just a Q&A machine, then Codex App is a digital employee that can truly sit in front of a computer and work for you.
It has extreme autonomy and can break down your one-sentence goal into a specific action plan. When executing each step in sequence, it calls various computer functions like a real person until the task is completed and the results are delivered to you.
Specifically, it can:
- Deeply read and write local files: Directly view, modify, and organize folders and project directories on your computer.
- Search real-time information online: It has a built-in full-featured browser, so you can search, browse, and summarize web information within the App without switching out.
- Control software like a human: It has visual perception capabilities and can recognize complex software interfaces in real-time to operate various Apps on your computer.
- Generate multi-dimensional work results: Automatically generate documents, images, web pages, code, and other results, with support for real-time previews.
- Seamlessly connect to external tools: Connect directly to your Gmail, Notion, Google Drive, and other applications through plugins.
- Background automation: After setting up scheduled tasks, it can automatically perform repetitive work in the computer's background.
Compared to the ChatGPT we are familiar with, Codex App brings two disruptive changes:
- Localization: Codex App runs on your local computer and can directly read and write local files; ChatGPT is a cloud-based AI chat application, and you must manually upload files for it to process them.
- Execution Capability: Codex App can perfectly execute any work, taking action and delivering results directly on your computer based on your needs; ChatGPT can only leave answers in the chat box, leaving the execution step to you.
In summary, Codex is a super assistant that truly lives in your computer and can do everything ChatGPT cannot.
II. Codex App Installation
1. Installation
Open the Codex official website (https://developers.openai.com/codex/app) to download.
It supports Mac/Windows; just choose the corresponding installation package for your computer.
2. Login
Log in with your ChatGPT account. The free version also has a small amount of credit, but it is recommended to start with a Plus membership.
III. Core Function Analysis
1. Global Overview
Entering the Codex App, the interface layout is shown below, mainly including the left navigation bar and the middle workspace.
Important functions in the left navigation bar include:
- Top bar: Plugins include plugins and skills, mainly used to expand Codex's capabilities; Automations are used to execute scheduled tasks.
- Middle chat area: The Projects area is used to manage project-level sessions bound to local folders; the regular Chats area is used to store scattered, temporary daily conversation history.
- Bottom settings: Used for global basic configuration.
The middle main workspace is the core workbench where all your chats begin.
Let's break down the core functions of Codex App specifically so you can get started quickly after reading.
2. Projects & Chats
The Projects and Chats areas are located in the left column and are two concepts you must master after entering Codex App. They represent two ways to start a session.
Projects: Corresponds to a local folder. Codex can directly read and write files in that folder during a project session. Therefore, projects are more suitable for tasks with clear goals, such as developing a website.
How to start: Click the new project icon to the right of "projects." It supports creating a new folder or selecting an existing folder to establish a project.
It is recommended to keep the session structure clear, mainly including these two points:
- Keep local folders clear, with different folders corresponding to different themes.
- Only advance one direction per session under the same project. For example, when developing a website, open different sessions for different functional designs. Don't mix them together; the execution will be better, and it will be easier to look back later.
Regular Chats: Not bound to any local folder, more suitable for scattered Q&A and chat needs.
How to start: Click the new chat icon in the "chats" area as shown below.
3. Plugins and Skills
Click Plugins in the left column to install Plugins and Skills for Codex to expand its capabilities.
Plugins are installation packages that connect Codex to external tools. For example:
- Gmail plugin: Let Codex read and manage your emails.
- Google Drive plugin: Let Codex read and edit your Docs, Sheets, and Slides.
- Vercel plugin: Let Codex deploy your project directly online.
Skills are reusable task instructions that tell Codex what to do when encountering a certain type of task. For example, the built-in Image Gen skill allows Codex to directly generate or edit images.
4. Automation
Automation allows Codex to automatically execute tasks in the background according to the time you set. It is suitable for repetitive tasks that need to be executed regularly, such as daily email summaries, scheduled information monitoring, stock daily reports, etc.
Setting it up is simple. For example, tell Codex directly in the chat:
Create a new automation to check my Gmail every morning at 9 AM and summarize important emails into the current chat session.
5. Chat Area Conversation Functions
1️⃣ Plan mode: Formulate a detailed plan before execution, suitable for complex tasks.
2️⃣ Permission mode: Before each chat, you can select the permission level in the bottom left of the input box. There are three levels:
- Default: Limited to the current workspace. When Codex needs to cross boundaries (access the network or files outside the workspace), it pauses to ask for your confirmation.
- Auto-review: Same boundaries as Default, but boundary-crossing requests are automatically reviewed and decided by AI without manual confirmation, avoiding workflow interruptions.
- Full access: Removes all boundary restrictions. Codex can access any file and the network, executing directly without interruption.
For daily use, Auto-review is recommended as it is relatively safe and controllable without frequent interruptions.
3️⃣ Model and Reasoning Level Selection: Supports selection of different models and reasoning levels.
Recommended: GPT 5.5 Medium for daily tasks, GPT 5.5 Extra High for complex tasks.
4️⃣ Input Box Shortcuts: There are three symbols in the chat input box to quickly trigger different functions:
- "@": Supports selecting and calling installed plugins or referencing specific files as context.
- "/": Brings up the command menu to execute built-in commands; installed Skills will also appear in the list.
- "$": Explicitly calls a specific Skill, such as entering $imagegen to directly trigger image generation.
IV. Recommended Settings
Before you start using it officially, it is recommended to configure these 4 basic settings for a smooth experience.
Click the settings in the bottom left corner to enter the configuration page.
- In permission settings, check auto-review and full access to easily adjust permission modes at any time.
- Enable the "Prevent computer from sleeping while Codex is running" option so that Codex's tasks are not interrupted.
- Configure Global Instructions (Agents.md)
In the custom instructions of personalization settings, write down the background information and rules you want Codex to follow long-term.
Recommendations include:
- Your identity and preferences: Such as your daily work role and commonly used software tools.
- Output format requirements: Such as "always answer in Chinese" or "keep the writing style concise."
- Principles of action: Such as "avoid over-design, only make necessary changes."
- Enable the Memory function in personalization settings. Codex will automatically learn your preferences from conversations and understand you better as you use it.
V. Super Practical Tips
1. Learn to use Plan Mode well
The more complex the task, the more it's worth opening Plan Mode to confirm the plan before execution. It's perfect for:
- Tasks involving changes to multiple files
- Requirements that aren't fully thought out yet
- Developing a function or Skill from scratch
2. Multi-task Parallel Execution
Codex supports starting multiple tasks simultaneously, even advancing across projects in parallel. The status of each chat is displayed in real-time in the sidebar, and a blue dot appears when a task is completed.
3. 4 Treasure Commands
- /side: While the main thread of a project-level session is running, open a temporary side chat for quick questions or confirming details.
- /pet: Open a desktop pet on the computer screen to display the execution status of the current task in real-time, allowing you to keep track of progress while doing other things.
- /status: Keep track of your context and credit usage in the session.
- /goal: Set a clear goal and let the agent work toward it.
A good /goal prompt looks like this:
/goal Achieve <the goal you want Codex to complete>, and confirm the result is valid through <specific verifiable evidence>, while maintaining <constraints that must be followed>. Only use <allowed inputs, tools, file ranges, or operational boundaries>. Between each iteration, Codex needs to choose the next step based on <how to determine the optimal next action>. If blocked or no valid path remains, Codex must stop and report <methods tried, evidence obtained, current blockage, and what information or permissions are needed to proceed>.
4. Steer Insertion Mechanism
While Codex is executing a task, enter a new instruction and click Steer (or Mac shortcut Cmd + Enter). The instruction will be immediately inserted into the current task's context without interrupting the ongoing work. This is suitable for making fine-tuning adjustments mid-run.
