Direct answer
If you are asking what is Codex, the current answer is OpenAI's coding agent. It can inspect a repository, edit files, run commands and tests, review changes, and return evidence about the work. You can use Codex from a terminal, supported code editors, the ChatGPT desktop experience, or a configured cloud environment. In current product language, Codex means this agent and its tools, not only the older OpenAI Codex language-model family.
Summary
Codex combines a coding model with repository context, tools, permissions, and a working environment. Give it a bounded task, inspect its actions, run the project's checks, and review the final diff. It accelerates software work; it does not transfer engineering accountability to the agent.
What is Codex in ChatGPT today?
Within OpenAI's current product family, Codex is the software-engineering agent. A model proposes the next action, while the agent supplies repository context and controlled tools. Those tools can read and change files, call the terminal, search when allowed, and use connected services. The surrounding product records the conversation, commands, diffs, approvals, and results so a person can steer and review the task.
The agent and the older model are different
The phrase Codex artificial intelligence can refer to two related subjects. OpenAI previously used Codex for a family of language models trained for code. The current Codex product is an AI agent that can use models and tools to complete software tasks. A model generates outputs; an agent works through a task inside defined permissions and returns artifacts for review.
How Codex works
- Goal: you describe a concrete outcome, constraints, relevant files, and how success will be checked.
- Context: Codex reads permitted project files, instructions, conversation history, and connected sources needed for the task.
- Plan: the agent breaks larger work into inspect, change, verify, and review steps when planning adds value.
- Action: it edits files or runs approved commands inside the active sandbox and permission policy.
- Feedback: command output, test failures, diffs, browser results, and your steering change the next action.
- Handoff: Codex reports the result, remaining risks, verification evidence, and the files or changes that need review.
Where you can use OpenAI Codex
- Codex CLI
- work against a local repository from the terminal, interactively or through repeatable codex exec commands.
- IDE extension
- keep the agent beside the code in supported editors such as VS Code, Cursor, or Windsurf.
- ChatGPT desktop
- coordinate local projects and longer tasks in the desktop experience described by the current docs.
- Codex cloud
- submit tasks to configured cloud environments and bring reviewed results back to the local workflow.
- SDK and integrations
- embed agent workflows or connect approved services through plugins and Model Context Protocol servers.
What Codex can do
- Explain an unfamiliar codebase, trace data flow, locate ownership, and answer focused architecture questions.
- Implement bounded features, refactor code, update documentation, and create tests while following repository instructions.
- Reproduce bugs, inspect logs, form a diagnosis, apply an authorized fix, and rerun relevant checks.
- Review uncommitted work, a commit, or a branch and report prioritized findings without changing the tree.
- Use images, current web context, skills, plugins, and MCP tools when the task and permission policy allow them.
- Delegate focused investigations to subagents and combine their results when parallel work is justified.
What Codex does not guarantee
Codex can produce incorrect code, incomplete tests, unsafe assumptions, stale facts, or a plausible explanation that misses the real cause. A passing command may cover only part of the product. A clean diff may still violate user needs, privacy rules, accessibility requirements, or deployment constraints. Treat agent output as work to verify, not an automatic approval.
Context and instructions shape the result
Codex performs better when it can find the repository's source of truth. An AGENTS.md file can describe commands, conventions, boundaries, and checks. More specific nested instructions can apply near the files they govern. Keep instructions current and concise. Link to existing documentation instead of copying rules into several places that can drift.
Permissions set the working boundary
A Codex session has an active sandbox, writable roots, network policy, and approval rules. These controls determine which files the agent can change and which commands need permission. Review them before a sensitive task. Give the narrowest access that still permits useful work, inspect risky commands, and keep publishing, billing, deletion, and external communication human-controlled unless explicitly authorized.
A safe first Codex task
- Open a small repository with a clean, recoverable Git state and read its contribution instructions.
- Ask Codex to explain one bounded area without editing files. Check the explanation against the code.
- Request a small change with acceptance criteria, file scope, constraints, and the exact verification command.
- Watch commands and diffs as they appear. Correct a wrong assumption before it spreads through the change.
- Run the relevant tests, type checks, lint, and build. Add targeted manual or visual verification when needed.
- Review the final diff yourself, confirm no unrelated changes, and record any residual risk before committing.
Frequently asked questions
ChatGPT Codex is the Codex coding-agent experience available through current ChatGPT plans and surfaces. It connects models to repositories, tools, working environments, permissions, and review records for software tasks.
No. The Codex name now identifies a coding agent and product, and it also referred to an earlier model family. Codex works with many programming languages rather than defining one.
Codex can automate and accelerate parts of software work. A person or accountable team still defines the problem, approves access, evaluates tradeoffs, verifies behavior, and owns the shipped result.