Getting started with Codex

Quick Overview

The Codex workflow is introduced through a session covering an overview, setup, prompting techniques, advanced use cases like MCP integration, and code review capabilities, all designed to streamline developer tasks from planning to release.

Key Points: Codex is OpenAI's coding agent, powered by GPT-5.1-Codex-Max, which runs locally or in the cloud via CLI or IDE extensions. Setup involves installing the Codex CLI via npm or brew, and authentication is done via a ChatGPT account using . Agents.md files provide project context and instructions to AI coding agents, with global configurations in /.codex/agents.md and project-specific ones in the local directory. MCP (Model Context Protocol) allows Codex to connect to external tools like Figma, Jira, Datadog, and Context7 for broader capabilities like fetching external documentation. Advanced prompting techniques include using to find issues against a base branch or uncommitted changes, and using images (via Ctrl+V or @file.png) in prompts. Codex supports headless mode for automation tasks such as Security Triage, Test Coverage Bot, Refactor & Cleanup Automation, and Release Hygiene Automation. The presentation concludes by pointing users to OpenAI resources like documentation, cookbooks, and the changelog for further exploration.

Context: This tutorial presentation, 'Getting started with Codex,' delivered by Derek and Charlie from OpenAI, introduces Codex, an AI coding agent derived from GPT-5.1-Codex-Max, detailing its architecture, setup process, fundamental prompting strategies, and advanced functionalities like integrating with external tools via the Model Context Protocol (MCP) and utilizing headless mode for CI/CD automation.

Detailed Analysis

The presentation introduces Codex, OpenAI's coding agent based on GPT-5.1-Codex-Max, emphasizing its ability to run locally or in the cloud via CLI or IDE extensions for fast iteration (00:16). The setup process requires installing the CLI (via npm or brew) and authenticating using a ChatGPT account via (05:58). A key concept is the file, which serves as a dedicated README for agents, providing project context and instructions, with global configurations stored in (09:20). The session moves into advanced capabilities, highlighting the Model Context Protocol (MCP) for connecting Codex to external tools like Figma, Jira, Datadog, and Context7 to leverage external documentation and context (03:37, 03:50, 03:55). The presenters demonstrate how to add an MCP server, like the Cupcake server, using (04:32). They also cover advanced prompting, such as using in the CLI to check uncommitted changes against a base branch (07:57, 43:18) and using images in prompts via Ctrl+V or @file.png, as shown by the example of adding buttons to the Agents.md site (19:18). Finally, the presentation touches upon Headless Mode for automating workflows like security triage, test coverage generation, refactoring, and release hygiene (49:02), demonstrating how Codex can be used to analyze code quality and output JSON schemas (04:22). The session concludes by pointing users to documentation, cookbooks, and the changelog for further learning.