Codex is OpenAI's software engineering agent, available as a command line tool, an editor extension, and a cloud service that works on tasks in its own sandboxed environment. Given a task it explores the repository, writes code, runs tests and returns a diff or opens a pull request. The cloud mode is designed for delegating several tasks in parallel and reviewing results later, while the local modes keep work on your machine with approval prompts before commands run. It integrates with GitHub for code review and can be driven from chat interfaces. It suits developers already paying for an OpenAI subscription and teams that want tasks handled asynchronously. The limitations are environment setup and verification: cloud tasks need their dependencies configured to be useful, results vary with how precisely the task was specified, and output still requires the same review as any other contribution.
How to use OpenAI Codex
1Install the command line tool through your package manager, run it in a project, and sign in with a paid OpenAI account; usage is included in subscription tiers with additional capacity available through the API.
2Start locally rather than in the cloud, with a contained task such as fixing a failing test, so you can watch how it proposes commands and approve them.
3Choose the approval mode deliberately: fully automatic execution is convenient and risky, and read-only mode is the safe way to evaluate.
4For cloud tasks, connect the repository and define the environment setup, including dependency installation, because a task that cannot install or run tests will produce plausible but unverified code.
5Write task descriptions the way you would write a ticket, with acceptance criteria, since precision matters more here than in interactive chat.
6Expect to review every diff, and keep tasks small enough that a rejected result costs minutes rather than a day.
Best for
delegated tasks
test fixing
parallel work
pull request generation
sandboxed execution
Pros and cons
What it does well
Runs locally or in a sandboxed cloud environment.
Handles several delegated tasks in parallel.
Integrates with GitHub for review and pull requests.
Included with existing paid subscriptions.
Where it falls short
Cloud tasks need careful environment configuration.
Results depend heavily on how precisely tasks are written.
Automatic approval modes carry real risk.
OpenAI Codex pricing
PaidPrices as listed by the maker; check the site for the latest.
Included in paid ChatGPT subscription tiers, with additional usage billed through the API