Do Claude skills work in Codex?

Often yes at the file level — both read SKILL.md. But each client expects skills in its own directory, and a skill tuned for one agent may need small edits to behave the same way in another. Test before you rely on it.

Which skills catalogue is authoritative?

The official repos — OpenAI’s skills catalog for Codex and Anthropic’s public skills repo — are the reference implementations. For third-party skills, provenance matters: prefer listings that show who published a skill, with ratings and sales rather than raw counts.

Are Codex and Claude skills free?

Official catalogues are free to browse and install. Community marketplaces mix free, free-trial and paid skills — on Capafy AI every price is shown up front, and several agents cost nothing at all.

Can I sell a skill to both audiences?

Yes. The Agent Skills format is shared, so one well-written skill can serve Codex and Claude users. On Capafy AI you publish once, buyers run it online, and you keep 80% of every sale.

Codex Skills vs Claude Skills — the Real Difference

The shared foundation: SKILL.md

Both Codex and Claude read skills in the Agent Skills format: a folder with a SKILL.md inside, optionally bundled with scripts, references and assets. The description in that file is the trigger — it tells the agent when the skill applies.

Because the format is plain text and the structure is identical, writing a skill is a transferable skill. If you can describe a workflow precisely, you can package it once and adjust it per client.

The two catalogues

Each ecosystem publishes its own reference catalogue. OpenAI maintains a Skills Catalog for Codex on GitHub; Anthropic maintains a public skills repository that anchors the Claude side. Around both, community marketplaces host thousands more skills — some listed per client, some cross-listed.

OpenAI — a documented skills catalog for Codex on GitHub
Anthropic — a public skills repo that the Claude ecosystem builds on
Community — directories and marketplaces indexing skills for both, often across 20+ clients

Where they actually differ

The same format does not mean the same experience. These are the differences that matter day to day:

The client — Codex lives in terminals and IDEs; Claude Code has its own environment, and the Claude apps are separate surfaces
The skills directory — each client loads from its own folder (Codex: ~/.codex/skills; Claude Code: .claude/skills or ~/.claude/skills)
Catalogue culture — OpenAI documents versions and bundles; Anthropic’s repo is the canonical starting point for Claude-native skills
The executing model — a skill is instructions, so behaviour can differ per model; test a ported skill before rolling it out

Why portability mostly works

A skill is text, and text copies. Moving one between clients usually means: place the folder in the other skills directory, tighten the description if the client activates it too often or too rarely, and re-test the tasks it covers.

The failure mode is not the format; it is assumption. A skill written against one model’s habits may need small edits for another — the steps hold, the wording changes.

One skill, every client

The marketplace route takes folders out of the equation. On Capafy AI, skill-based agents are packaged, published and rated — you run them online in the browser, no matter which client you prefer, and nothing needs installing.

Publishers keep 80% of every sale, and buyers get the same signal everywhere: real sales counts, ratings and a price up front. For a skill that works in Codex, in Claude and in five other clients, that is the shortest path from file tree to revenue.

Updated October 2026. Facts checked against the live Capafy AI marketplace.