Getting started with Codex

Codex is OpenAI’s coding agent that developers can delegate routine, time-consuming work to—freeing them to focus on design, architecture, and other novel engineering decisions. Teams are using it for practical workflows like code review before pull requests merge, Slack-thread-driven changes that produce new PRs, and custom integrations via an SDK that can run Codex inside their own containers. Codex also supports parallel background work through cloud environments, enabling async tasks such as code review while developers are away.

A key theme throughout the onboarding is that Codex performance depends on context management and guardrails. Codex sessions don’t retain project knowledge between runs, so agents.md acts like a lightweight, repo-specific “readme” that gets automatically loaded whenever Codex starts in a directory. It provides the agent with a project overview, structure, build and test commands, useful CLI tools, and workflow guidance for implementing features end-to-end. For larger efforts, teams can keep agents.md brief (often under 100 lines) and link out to task-specific markdown files—like plans.md templates that let Codex generate and maintain living execution plans for multi-hour refactors.

Getting started centers on installation and authentication. Codex CLI can be installed via Homebrew (brew) or npm, with npm recommended for staying current because releases ship quickly. The VS Code extension is installed by searching for “OpenAI Codex” and enabling auto-update. Access uses a Chatbot enterprise account: the CLI supports sign-in via “Codex login,” typically routing through internal SSO. Once authenticated, “/status” shows the active model, directory, sandbox mode, approval policy, and remaining context window.

Customization happens through config.toml, which controls defaults such as the model, reasoning effort, sandbox mode, approval policy, and optional features like web search. By default, approval is set to request escalated permissions only when needed, while sandbox mode limits file writes to the current workspace. Prompting best practices emphasize anchoring the agent to the right files (using file mentions), starting with small tasks, and including verification steps such as running tests or linters. For debugging, pasting full stack traces helps Codex navigate directly to the source of errors.

On the tooling side, Codex works across surfaces: a terminal-focused CLI for lightweight interactions and a VS Code-based IDE extension for richer navigation and workflows. Developers can bind keyboard shortcuts to add context, use IDE “to-dos” to queue tasks inside code, and even drive changes from images—like screenshot-based instructions to modify UI elements. Sessions can be resumed with “Codex resume,” preserving prior context as a mini project container.

The advanced portion highlights programmatic and enterprise-ready patterns. Codex CLI supports headless execution (“Codex exec”) with structured JSON output that can be parsed and fed into CI/CD pipelines for tasks like code-quality scoring, security triage, and automated refactor checks. For tool augmentation, Codex integrates with MCP servers over standard IO or HTTP—examples include Jira, Linear, Figma, Datadog, and a demo “Cupcake MCP.” It can also pull up-to-date framework documentation via context 7, reducing issues caused by documentation drift.

Finally, Codex includes built-in code review flows (CLI and IDE) that focus on high-severity findings (P0/P1) to avoid noisy feedback. For teams that want deeper automation, options extend to on-prem code review, CI autofix loops, and issue auto-triage by intent-based labeling. The session closes with pointers to developers.openai.com/codex, cookbooks, change logs, and admin/security/rate-card resources for enterprise planning and governance.