OpenAI Codex Beginner's Guide — Getting Started with Cloud-Based Coding Agent [2026]

OpenAI Codex is a cloud-based autonomous coding agent announced in May 2025. You give it a task in plain text, and it writes code, runs tests, and returns a diff — all inside an isolated cloud sandbox. Important: This is completely different from the original Codex (a GPT-3-based code completion model released in 2021). The current Codex is an autonomous agent built on o3/GPT-5 series models. Its purpose is task delegation, not line-by-line code completion. In February 2026, native desktop apps for macOS and Windows and IDE extensions were also released.

Here is a visual breakdown of how Codex processes a task. ``` [ChatGPT / Desktop App / CLI] ↓ [App Server] ↓ [Cloud Sandbox (isolated)] ↓ [Repository Clone] ↓ [Setup: Install dependencies (network ON)] ↓ [Agent: Edit code, run tests (network OFF)] ↓ [Diff + Logs + Test Results] ↓ [User Review → GitHub PR] ``` During the setup phase, internet access is allowed so npm, pip, or other package managers can install dependencies. The agent execution phase runs offline to prevent unintended external access.

Available interfaces

Model comparison as of April 2026

SWE-bench measures the percentage of real GitHub issues a model can autonomously resolve.

Follow these steps to get started with Codex immediately. ``` ① Subscribe to ChatGPT Plus or higher ↓ ② Open chatgpt.com or the desktop app ↓ ③ Open the Codex panel and connect your GitHub account (OAuth) ↓ ④ Select a repository and enter your task in plain text ↓ ⑤ Review the diff and test results, then create and merge a PR ``` Step ③ note: Codex requests read/write access to your repositories. During the first setup you can limit the scope to specific repositories for added security.

Codex offers two distinct modes.

Code mode spins up a full sandbox and actually writes code, so it takes longer — but you can work on something else while it runs. Ask mode reads the repository and answers in real time, ideal for quick clarifications. Recommended flow: Use Ask to design → Use Code to implement → Use Ask to clarify the diff → Merge the PR.

Supported languages (any language whose test runner works on a Linux container)

The 3 sandbox permission modes

Having automated tests dramatically improves output quality since Codex can self-verify its changes.

Using Codex requires a ChatGPT Plus plan or higher.

For large-scale enterprise use, Team or Enterprise plans are recommended. Direct API access is available on a pay-per-use basis, starting at approximately $1.50 USD per 1M input tokens depending on the model.

Q1. Does Codex work with private repositories? Yes. After connecting via GitHub OAuth, you can select any private repository. Code is deleted from the sandbox after the task completes. Q2. Is it compatible with the old Codex API? No. The original Codex models (e.g., davinci-codex) were deprecated in March 2023. The current Codex agent is accessed through the ChatGPT UI or a separate API endpoint. Q3. Does Codex automatically merge pull requests? No. Codex creates a diff and a draft PR, but merging always requires your approval. Q4. Can I use Codex without an internet connection? Codex is a cloud service, so an internet connection is required. There is currently no on-premises version. Q5. How large a codebase can Codex handle in one task? GPT-5.4 supports a 1M-token context window. Limiting each task to a focused set of files yields the most reliable results. Q6. Can Codex be used on a repository with no tests? Yes, but without automated tests Codex cannot self-verify its changes, so you will need to review the output manually. Q7. How is Codex different from GitHub Copilot and Cursor? Copilot provides real-time inline suggestions as you type. Cursor is an in-editor local agent. Codex autonomously executes complete tasks in a cloud sandbox and opens pull requests. Think of it as a delegated agent, not a completion tool.

Struggling with repository connection? Not sure how to break tasks down for Codex? Want to design a secure, enterprise-grade Codex workflow? Oflight's AI consulting service is here to help. We provide end-to-end support for adopting Codex and other cutting-edge AI agents — from implementation and process improvement to team training. Start with a free consultation. → View AI Consulting Services