Customizing Codex by letting the agent patch itself

A developer's X post argues that OpenAI's Codex agent can be used to modify the Codex desktop app itself, by reading the app's shipped code and inserting new features into it. The author says agents are "truly amazing at reading minified and compiled code", meaning code that has been compressed or converted for the machine rather than written for people to read, "and injecting their own code into it." The author says they built the Codex Subscription Router this way in an hour, noting that edge cases, such as how it handled resets, are normally what takes longest.

The method has three steps. Work from the Codex CLI, the command-line version, rather than the desktop app, because the agent can then restart the app without shutting itself down. Ask for a feature, such as making side chats persist across app restarts, keeping the usage limit visible next to the user's name, or giving projects theme colors. After each app update, ask the CLI to re-add the change from a mods folder, which the author says takes a minute. Each customization is therefore a patch that has to be reapplied, not a setting the app supports.

The author is open about the risks. Asked whether OpenAI will keep allowing this, the author answers "Maybe. Maybe not." The author encourages sharing mods but says to have Codex run a full security audit before installing one, because installed code gets read and write access to the computer. The post says the repository includes fixes for Codex's computer use and Appshots features, but that a user's own mods will break them, and suggests asking Codex to repeat the repository's fixes and check they work.

This is one developer's account, and the linked repository was not collected, so what the router does and whether it works as described are not verified here. The one-hour figure is the author's own.