OpenAI Codex: Skills, Plugin Tools, and the Orchestrator Agent

What is openai/codex, actually

openai/codex is OpenAI's own lightweight coding agent, built to run in a terminal rather than inside a browser or an IDE. The repository has passed 117,990 GitHub stars, and Flowy files it under the development category.

Flowy indexes the repository for what it ships alongside the agent itself: a small set of skills and one agent definition that shape how Codex behaves and what side tasks it hands off. In total that's 7 components: 6 skills and 1 agent. Read this guide alongside the listing page linked at the top, which has the real way to add it to your setup.

The skill that explains Codex to itself

openai-docs is the self-referential skill in the set. Its description points it at Codex's own models, pricing, scheduled tasks, skills, settings, setup, troubleshooting, customization, and automations, plus questions that use words like "you," "your," "this app," or "this coding agent" when the person actually means Codex itself. It also covers OpenAI's own APIs, products, and ChatGPT Work, among other things its description names. If someone asks the agent what models it supports or how to change its settings, this is the skill written to answer that instead of the agent guessing.

Three skills for building and installing more skills

Three of the six skills form a small pipeline for extending Codex with more of its own skills:

• skill-creator writes or updates a Codex skill, scoping its instructions to the task and pulling in whatever supporting files that skill needs.
• plugin-creator scaffolds a plugin directory around a skill: a required plugin.json manifest inside a .codex-plugin folder, optional plugin folders and files, working manifest defaults, and by default an entry in the user's own personal marketplace listing.
• skill-installer puts a finished skill into $CODEX_HOME/skills, either from a preset list bundled with the skill or by pointing it at a GitHub repo path, including private repositories.

Together they cover the whole loop: write the skill, wrap it as a plugin, then install it, without leaving the terminal.

Two skills for different jobs: images and review

The other two skills have nothing to do with skill-building:

• imagegen generates or edits raster images, for tasks where a bitmap visual, a photo, an illustration, a texture, a sprite, a mockup, or a transparent-background cutout, is genuinely the deliverable. Its description names creating a brand new image and transforming an existing one as two concrete triggers.
• review-agent runs a read-only, defect-first review of a specified code change and returns every actionable finding. It is written to be called by another agent, reviewing uncommitted changes, a diff against a base branch, a single commit, or a custom set of review instructions, rather than to be invoked directly by a person.

What does the orchestrator agent do

orchestrator is the plugin's one agent, and its description reads like a short list of ground rules for how Codex behaves. Two of them stand out. When a request is simple enough to satisfy with a terminal command, such as asking what time it is, it runs that command directly instead of routing the question anywhere else. And it treats the person it's working with as an equal co-builder, keeping their intent and their coding style intact rather than rewriting what they wrote. The description continues past that point, but the fact pack Flowy captured stops there.

Is this built for Claude Code?

No. Every skill here, and the orchestrator agent, describe Codex, OpenAI's own terminal coding agent, not Claude Code. Flowy indexes public GitHub repositories that ship skills and agents, and this repository is one of them. This guide exists so you can read what each piece does before deciding whether it's useful for your own setup, and the listing page linked at the top has the real way it gets installed.