How to Use OpenAI Codex: Core Concepts for Non-Coders Explained
A clear guide to installing Codex and understanding projects, agents.md, and agent loops for building AI workflows without writing code.
What is OpenAI Codex and how do you install it?
Codex is OpenAI’s agentic coding and automation tool, available as a desktop app, an IDE extension, or a command line interface. To install it, search for “Codex install” or find it through ChatGPT, since Codex is a product of OpenAI and ships alongside the ChatGPT desktop app. Download the version for your operating system and run the installer. If you already pay for a ChatGPT subscription, that same login works for Codex. No separate account is needed.
TL;DR
- Codex runs on your existing ChatGPT subscription, so if you already pay for ChatGPT there’s no new account to set up, though a free plan won’t be enough for real usage.
- Projects keep Codex organized around your work by grouping folders and files so the tool gets progressively more familiar with your business instead of starting from zero every session.
- The agents.md file sets ground rules for a project, acting like a briefing document that Codex reads before every message so you don’t have to reexplain context each time.
- The agent loop is the mechanism behind every response: the model reasons, calls tools, checks results, and repeats until it has a real answer instead of a guess.
- Codex can do everything a general “work” tool does and more, which makes it worth learning even for people with no coding background.
- Subscription pricing scales with usage, and heavier plans open up significantly more inference before you hit a weekly cap or need to pay per-use API billing.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does Codex differ from ChatGPT’s “work” mode?
OpenAI markets a “work” experience inside ChatGPT aimed at non-technical knowledge workers. Codex overlaps with a lot of that functionality but extends further, since it’s built as a full agent harness with file access, command execution, and tool use. The practical argument for choosing Codex from the start is simple: it can do everything the work-focused version does, plus more, so learning Codex first means you’re never boxed in in the future. If you eventually need capabilities beyond simple chat, you’ve already built the skill instead of having to relearn a new tool.
Neither option requires writing code. Both rely on natural language instructions. The difference is in ceiling, not in difficulty.
What are projects in Codex and why do they matter?
A project in Codex is just a collection of folders and files on your computer that Codex treats as a working environment. Rather than starting a fresh chat with no memory of your business every time, you point Codex at a project folder, and everything inside it (documents, credentials, notes, prior work) becomes context the agent can draw on.
Over time, a project can hold rules about who you are, ongoing initiatives, brand voice, and records of past conversations. Ask a well-built project for a summary of the last month’s work, and it can review its own history and files to answer accurately, rather than guessing. At that point the tool stops behaving like a simple chatbot and starts functioning more like a standing assistant with institutional memory of your operation.
What is an agents.md file and how do you use it?
The agents.md file is a plain markdown document that lives inside a project folder and defines the ground rules for how Codex should behave in that context. Despite the technical-sounding extension, it’s just a text file anyone can read and edit. It typically includes:
- A description of the agent’s role and priorities.
- Operating rules for how it should interact with you and your systems.
- Safety guidelines for actions it’s allowed to take on your computer.
- A “routing map” that tells Codex where to find specific information inside a large project (business documents, brand voice examples, active project files, and so on).
Every project can have its own agents.md, tailored to that project’s purpose. Before Codex processes any message you send, it reads this file first. That means the agents.md effectively primes every interaction, which is why the file matters more than any single prompt you type. A well-structured agents.md removes the need to reexplain context every session, since the rules and routing logic persist across every new chat inside that project.
What is the agent loop and why does watching it matter?
One coffee. One working app.
You bring the idea. Remy manages the project.
Codex, along with other agent harnesses like Claude Code, runs on something called an agent loop. At a basic level, when you ask a question, the agent doesn’t just generate a single text response. It reasons about what’s being asked, decides whether it needs a tool (like reading a file, searching memory, or pulling data from an external source), takes that action, evaluates the result, and repeats the cycle until it’s ready to respond.
You can actually watch this loop happen in the Codex interface. When a task runs, an expandable log shows each reasoning step: what the agent decided to check, which commands it ran, and what information it pulled before compiling an answer. For example, asking Codex to summarize the past month’s work triggers a sequence where it checks recent tasks, reads a memory file, scans previous chats, and only then produces a summary.
This transparency is genuinely useful for anyone learning the tool. Watching the loop in action teaches you how Codex thinks about a request, which helps you write better instructions in the future. If Codex consistently checks the wrong file or skips a step you expected, that’s a signal to adjust your agents.md or project structure, not just your prompt wording.
Is Codex worth the subscription cost?
Pricing follows the same tiers as ChatGPT: a free account isn’t sufficient for meaningful Codex use, and a paid plan (starting around $20 a month) is the realistic entry point. Heavier users on higher tiers, such as a $200 a month plan, get substantially more inference capacity before hitting a usage cap that resets weekly. Past that cap, pay-per-use API billing is available but priced at a premium compared to the subscription rate.
The framing worth keeping in mind: this isn’t an entertainment subscription. It’s closer to hiring capacity. For builders using Codex as an operating layer across content, business operations, and client work, the monthly cost is small relative to the output it can replace or accelerate, provided the person using it understands the core concepts (projects, agents.md, and the agent loop) well enough to direct it effectively.
How do you actually start using these concepts together?
The practical workflow looks like this: create a project folder for a specific area of your work, write an agents.md that explains your rules, priorities, and where key information lives, and then start interacting with Codex through natural language inside that project. As you work, Codex accumulates context, memory files build up, and each new session gets more useful because the agent isn’t starting cold.
None of this requires programming knowledge. The skill being built is closer to management: giving clear instructions, structuring information so an agent can find it, and reviewing its reasoning steps to catch mistakes early.
Frequently Asked Questions
Do I need coding experience to use Codex?
No. Codex is designed to be operated entirely through natural language instructions. Understanding core concepts like projects, agents.md files, and the agent loop is more important than any technical skill.
What’s the difference between a project and an agents.md file?
A project is the overall folder and file structure Codex works within, containing all your documents and history. The agents.md is a single file inside that project that sets specific rules, priorities, and routing instructions for how Codex should behave there.
Can I use Codex through something other than the desktop app?
Yes. Codex is also available as an IDE extension for tools like VS Code or Cursor, and through a command line interface, for people who prefer working in those environments instead of a standalone app.
Why does watching the agent loop matter for beginners?
Seeing each reasoning step and tool call Codex makes helps you understand how it interprets your requests. That insight makes it easier to write better agents.md rules and prompts, since you can see exactly where the agent looked for information and adjust accordingly.
Is Codex better than ChatGPT’s work-focused mode?
Codex covers the same ground as the work-oriented experience in ChatGPT and adds more capability on top, particularly around file access and tool use. Since both rely on natural language and neither requires coding, there’s little downside to starting with Codex directly.
