OpenAI Codex Skills Catalog Review: Reusable Agent Skills for Codex

If you have used Codex for more than a week, you have probably written the same deployment instructions or testing setup at least twice. The OpenAI Codex Skills Catalog tries to solve that — it is an official, open-source collection of packaged workflows that Codex can discover and run on demand.

This review breaks down what is in the catalog, how skills actually work, what it costs, and where it falls short compared to alternatives like Claude Code commands, Cursor Rules, and MCP servers.

What Is the Codex Skills Catalog?

A skill is a folder. At minimum it contains a SKILL.md file with YAML frontmatter (name + description) and markdown instructions. Optionally it includes scripts, reference docs, assets, and an agents/openai.yaml for UI metadata and MCP tool dependencies.

The catalog at github.com/openai/skills organizes skills into three tiers:

• System — bundled with Codex automatically (e.g. skill-creator, skill-installer)
• Curated — vetted by OpenAI, installable by name
• Experimental — community-contributed, installed via GitHub URL or folder path

As of March 2026, the repo has roughly 13,000 GitHub stars, 726 forks, and 35 curated skills.

How Skills Work

Codex uses progressive disclosure to stay context-efficient. It loads only skill metadata (name, description, file path) at session start. When a task matches a skill's description, Codex loads the full SKILL.md instructions on demand.

You can invoke skills two ways:

1. Explicitly — type $skill-name or use /skills to browse
2. Implicitly — describe your task and let Codex auto-select the matching skill

Installation is one command:

$skill-installer gh-address-comments

Restart Codex after installing. Skills load from four scopes: repository (.agents/skills in your project), user ($HOME/.agents/skills), admin (/etc/codex/skills), and system (bundled).

What Skills Are Available?

The 35 curated skills cover a wide range:

Deployment: Vercel, Netlify, Cloudflare, Render — each skill packages the platform's CLI setup, configuration patterns, and deployment commands into a repeatable workflow.

Design: Two Figma skills — one for general Figma integration and one specifically for implementing designs from Figma files into code.

Project management: Linear integration plus four Notion skills covering knowledge capture, meeting intelligence, research documentation, and spec-to-implementation workflows.

Testing & automation: Playwright (headless) and Playwright Interactive (with visual debugging powered by GPT-5.4's computer-use mode) for browser automation and testing.

Security: Three skills covering security best practices, ownership mapping, and threat modelling.

Content & media: Skills for slides, PDFs, spreadsheets, image generation, Sora video, speech synthesis, and transcription.

Developer workflow: gh-address-comments (resolve GitHub PR feedback), gh-fix-ci (diagnose and fix failing CI), and framework-specific skills for ASP.NET Core, WinUI, and Jupyter notebooks.

Creating Your Own Skills

You do not need to wait for OpenAI. The built-in $skill-creator wizard walks you through generating a new skill, or you can create one manually:

my-skill/
├── SKILL.md          # Required — frontmatter + instructions
├── scripts/          # Optional — deterministic behavior
├── references/       # Optional — docs, examples
├── assets/           # Optional — templates, images
└── agents/openai.yaml  # Optional — UI, policy, MCP deps

Commit skills to your project's .agents/skills folder and every team member's Codex instance discovers them automatically. This is the real power for teams — you package institutional knowledge into something Codex can execute reliably.

Pricing

The catalog itself is free and open-source. But running skills requires Codex, which has tiered pricing:

Plan	Price	What You Get
ChatGPT Plus	$20/mo	Codex CLI, IDE, web, iOS with usage limits
ChatGPT Pro	$200/mo	6x higher limits, priority processing
Business	$30/user/mo	Admin controls, SAML SSO, larger VMs
Enterprise	Custom	Audit logs, SCIM, zero data retention option
API Key	Usage-based	$1.50/M input tokens, $6/M output for codex-mini

There is no permanent free tier. ChatGPT Free and Go users get limited-time promotional Codex access, but it is not guaranteed to last.

Strengths

• Zero friction to start. $skill-installer plus a restart. No config files, no package managers, no build steps.
• Open standard. Skills follow the Agent Skills specification, published by Anthropic and adopted by OpenAI, Microsoft, Cursor, and others. Your skills are not locked to one vendor.
• Team knowledge packaging. Commit skills to a repo and your whole team gets them. No Slack messages, no wiki pages nobody reads.
• Progressive disclosure. Only loads what is needed, when it is needed. Context window stays clean.
• Active community. 94 pull requests, 77 commits, and growing. OpenAI's own team (Greg Brockman, Alexander Embiricos) actively promotes the catalog.

Limitations

• Codex-only in practice. The standard is open, but this catalog's installer and discovery mechanisms are Codex-specific. Moving skills to Claude Code or Cursor requires manual adaptation.
• No marketplace UI. Discovery is GitHub browsing or the CLI installer. No search, no ratings, no screenshots.
• No versioning. Skills have no semantic versions. There is no update mechanism — you reinstall or check the repo manually.
• Name conflicts. If two skills share a name, they appear separately in the selector. No merge, no priority rules.
• Implicit invocation quality varies. Auto-selection depends entirely on how well the skill's description field matches your prompt. No ML ranking layer.
• Restart required. After installing a skill, you must restart Codex. Minor but annoying during exploration.

How It Compares

vs Claude Code Commands

Claude Code uses CLAUDE.md files and /commands for project-specific instructions. Similar concept — structured markdown that the agent loads contextually. Claude Code's approach is more tightly integrated with its agentic workflow but has no central catalog or installer. If you are already in the Claude ecosystem, you build your own "skills" as commands.

vs Cursor Rules

Cursor's .cursorrules files serve a similar purpose inside the Cursor IDE. The main difference is scope — Cursor Rules are IDE-native and deeply integrated with Cursor's editor features, while Codex Skills are platform-level and work across CLI, IDE extension, and desktop app.

vs MCP Servers

MCP (Model Context Protocol) operates at the tool transport layer — it gives agents access to external tools and data sources. Skills operate at the instruction layer — they tell the agent how to do something. The two are complementary; skills can declare MCP dependencies in their openai.yaml config.

Who Should Use It?

You will get the most value from the skills catalog if you:

• Already use Codex CLI or the Codex app as your primary coding agent
• Work on a team where repeatable workflows (deployment, PR review, testing) matter
• Want to package domain knowledge into something an AI agent can execute
• Are comfortable with GitHub-based discovery (no marketplace UI)

If you use Claude Code or Cursor primarily, the catalog is less directly useful — though the underlying Agent Skills standard means the format is portable even if the installer is not.

Verdict

The OpenAI Codex Skills Catalog is a solid, practical extension system for Codex users. The 35 curated skills cover the most common developer workflows, installation is genuinely one-command, and the open standard gives confidence that skills are not a dead-end format. The biggest gap is discovery — no marketplace, no ratings, no visual browsing. And the Codex-only runtime means this is really for people already committed to the OpenAI ecosystem.