OpenAI Unveils Codex CLI, an Open-Source Agentic Coding Assistant That Can Operate Locally

OpenAI introduced an open‑source artificial intelligence (AI) command‑line tool on Thursday that can operate within a user’s terminal. Dubbed Codex CLI, it is essentially an AI agent that can perform various coding tasks and can even autonomously execute commands, all within a sandboxed local system. The tool does not require any installation setup, and can use multiple OpenAI AI models. However, the open-source project is currently under active development, and the AI firm warns that it might not be entirely stable.

OpenAI’s New Open-Source AI Agent Can Code Locally

The San Francisco-based AI firm introduced Codex CLI in a post on its website. Calling it a “lightweight coding agent,” the company said it can read, modify, and run code in a user’s terminal. The AI agent runs locally and does not require any Internet connectivity to function. It officially supports macOS and Linux, however, Windows support is experimental and might require Windows Subsystem for Linux (WSL).

OpenAI’s Codex CLI is an interesting project. Today, developers have two options if they want to use AI systems for coding-related tasks. They can either opt for a cloud-based closed AI system, which would require them to upload their code to the cloud and pay for the service. Or, they can opt for an open-source AI model that can be downloaded on their workstation, but it is generally less powerful compared to what companies offer.

Codex CLI offers a midpoint between these two solutions. OpenAI says the AI agent can operate locally, meaning coders do not have to share their repository with the AI firm. This ensures data security for the developers. On the other hand, the AI agent runs on the company’s powerful o4-mini and other reasoning models, offering flagship-tier performance.

“Because the CLI runs locally, your source code never leaves your environment unless you choose to share it,” the company said on its website. However, user prompts, high-level context, and optional differential summaries are sent to the model for output generation.

The AI agent does not require any installation setup, instead, it can be activated using an OpenAI application programming interface (API) key. By default, Codex CLI uses the o4-mini AI model, but the company is also letting users switch between multiple other options. Notably, since it requires an API key, the AI agent is not free, even if it can be run locally. API charges will apply.

There are three modes that developers can use. The Suggest mode can read files in the user’s repository, propose edits and shell commands, however, it will need permission before making changes or executing commands. In the Auto Edit mode, it can read and write files automatically, but it will still ask for permission before running shell commands.

Finally, in the Full Auto mode, the AI agent can read, write, and execute commands autonomously inside a sandboxed, network-disabled environment scoped to the current directory. Codex CLI can be accessed from OpenAI’s GitHub listing.