Best Codex CLI Skills in 2026 (Ranked, Free + Paid)

- Skill sources, repo stats, and archive status were checked directly against GitHub and the official docs at the time of writing; this space moves fast, so verify live before you depend on any of it.

At a glance

One row per pick: what it does and how to get it. Full detail on each is below.

What counts as a "Codex skill" here

A Codex skill is a SKILL.md folder Codex loads on demand when your request matches its description - not something always sitting in context like AGENTS.md. A minimal one is just a folder with a name and a short trigger description at the top; the actual instructions only enter context once Codex decides the description matches what you asked for, which is why a project can carry dozens of skills without bloating every prompt. This piece assumes you already know that and ranks picks; if you want the full mechanics (discovery order, folder scopes, how to write your own), read what a Codex skill actually is first.

One quick distinction worth knowing before the list: not everything below installs the same way. Most are true skills, installed via $skill-installer or a manual clone into .agents/skills. One - Superpowers - ships as a Codex plugin instead, installed through the separate /plugins interface. Same goal (extend what the agent knows how to do), different install path; see the FAQ for the full distinction.

How we picked these

Every repo below was checked live via GitHub - stars, last push, archived status, and the actual directory listing - on 2026-08-20, not pulled from an older write-up. That check caught two things worth flagging up front: mcp-builder is commonly cited as an openai/skills entry, but the live directory listing shows it doesn't live there - it's in anthropics/skills at skills/mcp-builder. And the skill some 2026 write-ups call "create-plan" is currently named define-goal in the live catalog. Small details, but they're exactly the kind that break a copy-pasted install command.

Two things got a skill cut from this list. First, a placeholder install command - one competitor roundup literally shows github.com/example/... as its install source, which is a pretty clear tell that the entry was never actually run. Second, an install-count claim with no visible public source (a few pages cite figures like "764 installs" with nothing backing them). Everything ranked below is something we could clone, run, or point $skill-installer at directly.

One more filter worth naming: a skill earns a slot here by doing something an agent genuinely struggles with unassisted - shelling into gh correctly, driving a real browser, scaffolding MCP boilerplate - not by wrapping a one-line shell command in a SKILL.md folder for the sake of having an entry. That's also why this list stops at seven individually-ranked picks instead of stretching to ten or twelve padded with marginal utility skills.

Best individual free Codex skills

1. Superpowers - a cross-tool development methodology, not a single-purpose utility: it walks an agent through brainstorming, planning, and test-driven implementation as separate, deliberate stages instead of letting it freestyle straight to code. The pitch is the same one behind define-goal below, just broader in scope - fewer confidently-wrong "done" claims, more of a forced pause between "I understood the ask" and "I'm writing code." Source: github.com/obra/superpowers, ~274k stars as of 2026-08-20 (verify before quoting - this moves fast). On Codex, its own README documents installing it as a plugin: run /plugins, search "superpowers," and select Install Plugin. That flow points to what the README calls "the official Codex plugin marketplace" at github.com/openai/plugins - the same repo that's now archived (see the heads-up section below), so treat that specific link as informational rather than a live browse-and-discover source.

2. mcp-builder - scaffolds a working MCP server (tool definitions, schema, boilerplate) from a plain-language spec, so you're not hand-writing MCP glue code from scratch every time you want Codex to talk to a new API or internal system. Worth it the moment you're wiring up a third or fourth custom MCP server and the boilerplate starts feeling repetitive. Source: github.com/anthropics/skills, path skills/mcp-builder (confirmed via live directory listing 2026-08-20 - not openai/skills, despite how a few competitor pages cite it). No confirmed cross-repo shortcut through $skill-installer, so clone it directly: git clone https://github.com/anthropics/skills /tmp/anthropic-skills, then cp -r /tmp/anthropic-skills/skills/mcp-builder .agents/skills/, then restart your Codex session.

3. skill-creator - the built-in meta-skill for authoring your own SKILL.md: it interactively walks you through the name, trigger description, and folder layout instead of you copying a template blind. Source: openai/skills, skills/.system/skill-creator - confirmed live in the directory listing 2026-08-20. Install: $skill-installer skill-creator.

4. gh-fix-ci - reads a failing GitHub Actions run and proposes (or applies) the fix, instead of you tailing logs by hand across five re-runs. Like Codex's own /review, expect it to shell out to gh for run status and logs - have the GitHub CLI installed and authenticated locally before you try it. Source: openai/skills, skills/.curated/gh-fix-ci, confirmed live 2026-08-20. Install: $skill-installer gh-fix-ci.

5. gh-address-comments - walks a PR's review comments one by one and applies the requested fixes, so "address CI feedback" stops meaning "re-read the whole thread manually and hope you didn't miss one." Same gh-CLI dependency as above. Source: openai/skills, skills/.curated/gh-address-comments, confirmed live 2026-08-20. Install: $skill-installer gh-address-comments.

6. playwright - terminal browser automation for real end-to-end testing and screenshot verification, so "it builds" and "it actually renders correctly" stop being the same claim. This is the tool that makes an agent-driven "verify before you say done" workflow possible instead of a trust exercise. Source: openai/skills, skills/.curated/playwright (a playwright-interactive variant lives in the same catalog if you want a step-through mode instead of fully autonomous runs). Install: $skill-installer playwright.

7. define-goal - turns a vague request into a scoped goal with an outcome, constraints, and a stop condition, before you hand anything to an autonomous loop; it's the difference between "make it better" and something an agent can actually execute against without drifting. Source: openai/skills, skills/.curated/define-goal - note this is the current live folder name; some older 2026 roundups reference the same idea as "create-plan," which no longer matches the catalog. Install: $skill-installer define-goal.

A pattern across five of these seven: they're narrow on purpose. gh-fix-ci only fixes CI, gh-address-comments only handles review threads, playwright only drives a browser. That's the actual design goal of a skill - one job, loaded only when needed - versus a general-purpose instruction sitting in AGENTS.md that Codex re-reads on every single turn whether it's relevant or not.

Best places to browse for more skills

Directory counts and star totals here move by the day, and two of these read almost implausibly high for what are, at heart, folders of markdown and shell scripts - that's just where this niche sits right now. Re-check any of these numbers before you quote them, especially the two sitting in six figures.

Which one to actually browse depends on what you're after. anthropics/skills is the safest general pick - it's where mcp-builder lives, and being Anthropic's own repo, it tends to get skills written carefully rather than dumped. mattpocock/skills reads as one practitioner's real, in-use folder rather than a curated showcase, which can mean more opinionated (and occasionally more useful) picks. vercel-labs/agent-skills leans toward frontend/deploy workflows given Vercel's own product surface. composio-community/awesome-codex-skills is the closest thing to a link-aggregator of the bunch - useful for discovery, less useful as a source you'd clone directly without reading first.

Heads up: openai/skills vs openai/plugins is confusing right now

Here's the state of it, plainly, with dates. openai/skills - the official Codex skill catalog, and the source behind four of the seven picks above - carries its own README notice saying it's deprecated and pointing readers to openai/plugins for "current Codex skill and plugin examples." That successor repo was archived by its owner on 2026-08-16 and is now read-only. So the deprecation notice sends you to a dead end. Meanwhile the current skill-building doc (learn.chatgpt.com/docs/build-skills, checked 2026-08-20) still cites openai/skills - not openai/plugins - as its example link. It's not just the docs, either: Superpowers' own install instructions point Codex users to what its README calls "the official Codex plugin marketplace" at github.com/openai/plugins for the /plugins search interface - the same now-archived repo.

Practical takeaway: openai/skills itself is not archived, so its skill folders are still there to browse and clone - that's exactly what powers most of the ranked list above. But don't treat either GitHub repo as a growing, actively-maintained catalog going forward; expect $skill-installer's default source, or the in-app Plugins surface, to keep shifting. Bookmark learn.chatgpt.com/docs/build-skills as the one link worth trusting long-term. Re-check both repos' status before you publish or automate anything against this - it can flip again.

Why this matters beyond trivia: if you're scripting an install step (a team onboarding doc, a Dockerfile, a setup script), hardcoding a git clone against either repo is now a fragility risk - one is already read-only and the other carries an active deprecation notice pointing at a dead end. Point automation at $skill-installer or the official docs instead of a pinned repo URL, and you won't have to revisit the script the next time one of these flips status.

Don't want to hunt or assemble your own? AgentKit (paid, one honest pick)

One disambiguation first: this is AgentKit at agentkit.best (CLI ak), not OpenAI's own AgentKit product. See OpenAI AgentKit vs AgentKit if that name is confusing you.

Every skill above is free. AgentKit is different: it's a paid, licensed bundle of skills, subagents, and workflows that installs into the same .agents/skills path Codex already reads, rather than one skill at a time. ak kit init engineer --target codex --global installs it; inside a Codex session you invoke it with $ak:cook (Claude Code uses the same kit via /ak:cook instead - one license, one runtime flag). It's a fit for people who'd rather buy a curated, maintained set - see AgentKit's Codex kit - than piece one together from the seven picks above and keep re-checking whether each one still resolves. Full walkthrough: using AgentKit inside Codex. Full write-up: the AgentKit review. Note this is a separate purchase from Codex itself - Codex stays free or bundled into whatever ChatGPT plan you already pay for either way.

Worth being blunt about the trade-off: assembling the seven free skills above costs you nothing but the time to check each source, install it, and notice when one of the repos above changes status. AgentKit costs money and buys you exactly that maintenance work, done once, with an install path that doesn't depend on you tracking which GitHub repo is deprecated this month.

How to install anything on this list

Two patterns cover everything above. For anything in the official catalog, run $skill-installer <name> from inside Codex (for example $skill-installer gh-fix-ci), and it clones straight into your skills folder - user scope by default, or pass a project path if you want it repo-scoped so the rest of your team gets it too. For anything outside that catalog - mcp-builder, or a plugin like Superpowers - clone the source repo and copy the specific folder into .agents/skills/ (repo scope, commit it so teammates get it) or $HOME/.agents/skills (personal, every project), then restart your Codex session so it picks up the new folder. Either way, skim the skill's SKILL.md and any bundled scripts before you trust it - the same review you'd give a new dependency. If a skill you just installed doesn't seem to trigger, the usual culprit is the trigger description being too narrow or too generic for what you actually typed - open the SKILL.md and check the description line before assuming the install failed. For the full discovery order, scopes, and how to write your own SKILL.md, see what a Codex skill actually is - this page ranks, that one teaches.

Frequently asked questions (FAQ)

Are Codex CLI skills free?

Yes - every individually-ranked skill above is free; SKILL.md is a native Codex feature already included in your ChatGPT plan. The only paid option on this list is AgentKit, a third-party curated bundle you can choose instead of assembling skills yourself.

What's the difference between a skill and a plugin in Codex?

A skill is a SKILL.md folder loaded on demand into a coding session, installed via $skill-installer or a manual clone into .agents/skills. A plugin installs through the separate /plugins marketplace interface - Superpowers, for example, ships as a plugin on Codex rather than a skill-installer skill. The two overlap in purpose (both extend what the agent can do) but use different install paths and, at least for now, different discovery surfaces.

Is openai/skills still the official skill catalog?

Yes, but with a catch. openai/skills is still live and not archived, but its own README flags it as deprecated in favor of openai/plugins - and openai/plugins was itself archived (read-only) by its owner on 2026-08-16. The official docs at learn.chatgpt.com/docs/build-skills still cite openai/skills as the example link, so it remains the more useful of the two to browse, even carrying a deprecation notice that currently points nowhere useful.

Do these skills work in Claude Code too?

The ones built on the open SKILL.md format do - Codex and Claude Code both read it, and mcp-builder, gh-fix-ci, and the rest install the same way in either tool. Superpowers explicitly documents a separate install path per harness, so check its Claude Code section rather than assuming the Codex steps carry over.

Is AgentKit an OpenAI product?

No. AgentKit here is agentkit.best, a third-party kit driven by the ak CLI. OpenAI's own AgentKit is a different, unrelated product (Agent Builder/ChatKit) - see our full disambiguation for details.

Which skill should I install first?

If you only try one, start with skill-creator so you can see how a real SKILL.md is structured, then add gh-fix-ci or gh-address-comments if you live in GitHub Actions and PR review. Superpowers and mcp-builder are worth it once you're ready for a bigger workflow change.

Conclusion

Start with two or three free skills - skill-creator plus whichever GitHub or MCP skill matches your actual workflow - before you go looking for a full catalog to browse. There's no official ranking of Codex skills, here or anywhere; this list reflects one team's picks after independently checking every source, not a definitive top-N. openai/skills is still the more useful of the two official repos to check, but treat both as a snapshot, not a living catalog, and bookmark the official docs instead. If you'd rather skip assembling any of this yourself, AgentKit is the one paid, honestly-labeled option on this list. For the broader picture beyond just skills, see the best AI coding tools roundup.