Working with OpenAI’s Codex CLI

A developer documented a detailed setup process for OpenAI's Codex CLI, a terminal-based tool for generating, refactoring, and documenting code with AI models. The guide covers configuring API keys for providers including OpenAI, Gemini, OpenRouter, Ollama, Mistral, DeepSeek, xAI, and Groq, creating a persistent config at ~/.codex/config.json, and adding a global ~/.codex/instructions.md for context passed to every command. The developer noted the CLI does not yet respect .gitignore or support standard ignore rules, though open pull requests aim to add both.

OpenAI’s Codex CLI is a developer command-line interface that lets you leverage AI models to generate, refactor, and document code directly from your terminal. This video provides a full demo and explanation, and I highly recommend watching it. It shows just how powerful the codex CLI is, they even create an SPA version of Mac’s Photo Booth app using an uploaded screenshot and short description. It’s pretty amazing.To start using it in your day-to-day workflow as a developer you’ll need to install and configure it. The guidelines in the readme.md aren’t all that great, so I’m documenting the steps I’ve taken here along with many undocumented aspects I learned along the way. To work with the Codex CLI we need to start with a clean setup. Here’s the process as I see it, which is more detailed than the official documentation but includes some critical in my mind required steps like setting up persistent configuration, adding default context instructions for all commands, and most important: adding ignore rules for input. First you’ll need to set up your API keys for the AI APIs you’ll be working with. As of today there is support for: openai default openrouter gemini ollama mistral deepseek xai groq Other providers that are compatible with the OpenAI API will also work, so there may be other options. I use Gemini and OpenAI, which is pretty common so if you’re planning to use these you can setup API keys at the following URLs: Once you have those you’ll need to add them to your command line profile so that they’re available to node’s process.env globals. I use Fish CLI https://fishshell.com , so for me this looks like this: Codex CLI set -x OPENAI API KEY {YOUR API KEY} set -x GEMINI API KEY {YOUR API KEY} It’s a good idea to verify that these are configured and available, so after you save them reload your CLI profile and then verify they’re good to go with the following test commands: echo $OPENAI API KEY node -e 'console.log process.env.OPENAI API KEY ' echo $GEMINI API KEY node -e 'console.log process.env.GEMINI API KEY ' If you see the correct values printed after running each of these then you’re good to go. After that we’ll create a config file at ~/.codex/config.json and adjust the settings for however you want to use it. Reference the official configuration guide //~/.codex/config.json provided in the openai/codex GitHub repository for all the options. I use either OpenAI’s gpt-4.1 or Gemini’s gemini-exp-1206 as of May 7th, 2025 . I’m sure this will change in the future. Here’s what my config looks like: { "model": "gpt-4.1", "provider": "openai", "reasoningEffort": "medium", "providers": { "openai": { "name": "OpenAI", "baseURL": "https://api.openai.com/v1", "envKey": "OPENAI API KEY" }, "gemini": { "name": "Gemini", "baseURL": "https://generativelanguage.googleapis.com/v1beta/openai", "envKey": "GEMINI API KEY" } }, "history": { "maxSize": 10000, "saveHistory": true, "sensitivePatterns": } } If you want to use Gemini then the top of the config would look like this: { "provider": "gemini", "model": "gemini-exp-1206", ... } These models will probably change and there will be better ones in the future, so be sure to reference the latest Gemini and OpenAI models and choose the best ones for your specific use case, budget, and expected outcomes. Right now there isn’t a standard way to define ignore rules for the codex CLI, and it doesn’t respect the .gitignore either. There are open pull-requests that add support for both of these features, hopefully in the near future we’ll see those merged into main . Add a ~/.codex/instructions.md file to provide the CLI with a global context that is used for all commands you run. This will be passed to every command you run and can include things like: - use comments minimally, especially when inline - adhere to any coding style guide configurations found - inline comments have an empty line above, and are on their own line It’s a new tool and I’m actively working with it. After a few weeks I plan to either update this portion of the article or write a new one with details on quirks and tips about how effective it is, and what will make using it easier. One of the most important aspects of the new CLI is the coding model, which has 3 different configurations: --suggest Generates code suggestions based on your input, but does not modify files automatically. You manually review and apply the suggestions you want. openai codex --suggest --auto-edit Automatically applies code changes to files based on your prompts, but asks for confirmation before each change. openai codex --auto-edit --full-auto Fully automates code changes without asking for confirmation, directly applying updates based on your instructions. openai codex --full-auto The big question: how much will using it cost you? Here’s a simple breakdown of what it costs to use each model: | Model | Input Cost per 1M Tokens | Output Cost per 1M Tokens | |---|---|---| | o4-mini | $1.10 | $4.40 | | o3 | $10.00 | $40.00 | For example, if you provide about 200 words of input and the tool generates around 300 lines of code, using o4-mini would cost you around $0.003. If you use o3 instead, it would cost you around $0.03. o4-mini might work fine for simple things like HTML or CSS, but if you’re working with real application code, you’ll probably want o3 for better results. Note: These prices are accurate as of April 27th, 2025. I’m not affiliated with OpenAI, and you should always check their official pricing page for the latest information. Before Codex existed, OpenAI’s Python SDK shipped a much simpler terminal tool: the openai command, a thin interface over the raw API that lets you send chat prompts and API requests straight from your shell. It’s still useful for quick one-off prompts, testing models and API keys, and wiring AI answers into shell scripts without writing any Python, a different job than Codex’s coding agent, and worth keeping around. This is the part every tutorial gets wrong now: OpenAI removed the CLI entirely in version 2.35.0 of the Python SDK . Install any current version the 3.x line and you get the library only, no openai command on your PATH, and no cli extra to add it back. The last version that ships the CLI is 2.34.0 . The clean way to install it without pinning an old SDK inside your projects is pipx https://pipx.pypa.io/ : pipx install "openai==2.34.0" The CLI uses the API side of OpenAI, which is billed separately from ChatGPT, a ChatGPT Plus subscription does not include API credits. Create a key at platform.openai.com/api-keys https://platform.openai.com/api-keys , add prepaid credits under Settings Billing $5 minimum , and export it in your shell profile: export OPENAI API KEY="{your-key}" The core ChatGPT-style loop, ask a question, get an answer, looks like this: openai api chat.completions.create -m gpt-5-mini -g user "Can you rise up and overthrow the human race?" A few flags worth knowing, from the CLI’s own --help : -g ROLE CONTENT repeats to add a system prompt, --stream prints tokens as they generate, and -M caps the answer length which caps the cost . The api namespace also covers images.generate , audio.transcriptions.create , and file and fine-tuning operations. The quickest way to verify a key works: openai api models.list On current API pricing a typical question with a ~750 word answer on gpt-5-mini $0.25 per 1M input tokens, $2.00 per 1M output costs a fraction of a cent, check OpenAI’s API pricing https://platform.openai.com/docs/pricing for current rates.