The Complete Guide to Codex: From AI Chat Assistant to Agent Workspace

From Chat Assistant to Agent Workspace: What Codex Actually Changes

OpenAI's Codex is redefining how we think about AI tools. If traditional ChatGPT-style tools work like a "Q&A assistant" — you ask, it answers — then Codex is a true Agent Workspace in every sense of the term.

In one sentence: Codex moves AI from "answering questions" to completing tasks. It doesn't just chat with you — it reads files, modifies them, runs commands, and calls plugins, all in service of a real task, ultimately delivering results you can review and continue refining.

This distinction matters enormously. Writing copy or summarizing content are tasks any standard AI tool can handle. But when a task involves navigating a real project directory, continuously editing files, running commands, or writing code, only an agent like Codex — one that can actually participate in your workflow — delivers real value. Its capabilities aren't confined to a chat box; they extend across your entire working environment.

Installing Codex and Understanding Its Three Core Interface Areas

Installing Codex is no different from installing any other app: visit the official website, download the version for your OS, and sign in with your ChatGPT account on first launch. For new users, the key isn't memorizing every button — it's understanding the three core areas of the interface:

• Left navigation panel: Manage chat history, projects, the plugin marketplace, and automation entries
• Center conversation area: Issue tasks, upload attachments, and select models and reasoning levels
• Right results panel (workspace): Display file previews and generated outputs

A great starter exercise: send Codex the instruction "Create a Markdown file explaining the difference between Codex and a regular chat tool in three sentences." You'll notice it doesn't just reply in the chat window — it actually creates a file in your project directory.

This seemingly simple action is the foundation for everything Codex can do — it has a real workspace where it can read and write files and execute operations.

Chat vs. Project: Two Work Modes Explained

Codex offers two core ways of working. Understanding the difference is essential to using it effectively.

Chat Mode: For Quick, One-Off Conversations

Chat mode is closer to a regular chat experience, suited for quick questions. Ask something like "explain what MCP is" and Codex responds like a standard AI tool. These interactions don't require long-term file management and don't produce persistent local results.

Project Mode: For Persistent File-Based Work

Project mode is designed for serious tasks. Whenever a task will produce files — building a webpage, editing a presentation, working with spreadsheets, writing code, reading source material, making ongoing edits, running commands, or calling plugins — Project mode is the right choice.

When creating a project, you can choose "Create a new blank project" (essentially creating a new folder) or "Use an existing folder." Each project maps to a real working directory, which you can right-click to rename or view the actual path. A simple rule of thumb: Chat is for temporary conversations; Project is for formal work that produces persistent results.

Model Selection, Reasoning Levels, and Permission Settings

These three settings control Codex's "brain," "thinking depth," and "freedom of action."

Model selection is the brain of Codex. In the bottom-right corner of the input area, you can choose from GPT-5.5, 5.4, 5.3, 5.2, and other versions. New users don't need to overthink this — the default model handles the vast majority of tasks just fine.

Reasoning levels come in four tiers: low, medium, high, and ultra-high. Higher levels take more time and are better suited to complex tasks like planning large-scale projects or analyzing large volumes of files. There's also a speed option — fast mode is 1.5× faster than the default, but consumes twice the tokens.

Permission management deserves the most careful attention. When executing tasks, Codex may read and write files, run commands, access the internet, install dependencies, and even control your browser or computer — all of which require authorization. An important note: more permissions aren't always better. The higher the permissions, the more freedom Codex has, but also the more important it is to keep an eye on what it's actually doing.

In practice, when an instruction asks Codex to create a file outside the current directory (such as in the Downloads folder), it will prompt for confirmation under default permissions. This shows that by default, it can only access the current directory. New users are advised to stick with default or auto-review permissions initially, and to read each prompt carefully before confirming — check what it wants to access and what command it plans to run.

Codex Plugin Marketplace: A Toolbox for Specialized Capabilities

Plugins is Codex's "toolbox" — installed plugins unlock handling of more specialized tasks. Take the most practical set, the Office Suite plugins, as an example:

• Documents plugin: Create and edit Word documents
• Spreadsheets plugin: Work with Excel and CSV files
• Presentations plugin: Generate and modify PowerPoint presentations

In a real-world example: give Codex an "AI tools learning resource" document and ask it to "organize this into three outputs: a Word study guide, an Excel tool comparison chart, and a PowerPoint course outline." It will automatically invoke the corresponding three plugins and generate all three file formats in one go. You can also manually invoke a specific plugin by clicking "+" and selecting it explicitly.

Plan Mode: Plan First, Then Execute — Avoid Costly Rework

For complex tasks, Codex offers Plan mode, which follows a four-step flow: understand → break down → confirm → execute.

Consider a typical scenario: you want to build a personal blog. If you let Codex jump straight in, the tech stack, code structure, and visual style it produces are likely to miss the mark. The point of Plan mode isn't to "slow things down" — it's to have Codex break the task apart first, tell you what it plans to do, and only proceed once you've confirmed.

In the blog example, Codex first presents a dialog asking you to choose the blog's purpose, implementation approach, and visual style, then lists a complete plan. Once you're satisfied, click "Implement this plan" and it gets to work. Plan mode is especially valuable when you have clear expectations for the outcome and don't want the AI to improvise.

Skills and MCP: Work Instructions and External Connectors

These two concepts are easy to confuse but critically important.

Skills: Operating Instructions for Codex

A Skill is like a step-by-step guide that tells Codex what steps to follow, what rules to observe, and what resources to use when performing a certain type of task. There are image-generation Skills, document-handling Skills, and many more. GitHub hosts a large collection of ready-made Skills covering development, writing, data analysis, and beyond.

Installation is straightforward: click to install recommended Skills from the "Skills" panel; copy a third-party Skill's install command from the terminal and run it with one click; or click "Skill Create" to build your own and describe your custom workflow.

MCP: The Standard Interface for Connecting External Tools

MCP (Model Context Protocol) is the standard interface for connecting external tools and services, enabling Codex to integrate with browser automation, databases, design tools, project management systems, and more.

You can configure MCP via the terminal command codex mcp add followed by the address, or simply send "Help me install Figma MCP" in the input box. With Figma configured, for example, Codex can read your Figma design files — bridging the gap from design to code.

One-line distinction: Skills teach Codex how to do something; MCP lets Codex connect to external systems.

Automations and Computer Use: Codex's Advanced Capabilities

Automations handle repetitive tasks. For example, you can set up a task like "every evening at 8 PM, fetch the latest Codex news from social platforms and generate a Markdown file." Codex will execute scheduled tasks automatically. You can monitor task status, check the next run time, and start, pause, or delete tasks from the Automations panel at any time.

Computer Use is Codex's most powerful capability — once authorized, it can directly operate your computer: clicking buttons in applications, typing text, taking screenshots, and reading on-screen information. In a demo scenario, Codex was even able to open a chess app on the computer and play a game against it. That said, this capability carries extremely high permissions, and new users should approach it with great caution.

The Right Mindset for Getting the Most Out of Codex

Truly mastering Codex isn't about memorizing every button — it's about adopting one core mindset: communicate the task clearly: what you want done, what reference materials to use, and how to verify the final result.

Once you internalize this approach, whether you're writing, doing office work, learning, or coding, Codex can save you enormous amounts of repetitive effort. It represents more than just another AI tool — it marks a paradigm shift in AI, from "conversation" to action.