Claude Code, Codex and OpenCode all load skills in the same format: a folder with a SKILL.md file inside. The format is the open Agent Skills standard, which Anthropic created and which became a Linux Foundation project on November 4, 2026. So one skill can work in all three tools. The catch is that each tool looks in different folders, and some fields only Claude Code understands. This guide shows where each tool looks, how to keep one copy that all three load, what doesn't carry over, and how Orkas gives the same skills to every coding agent it runs.
The short answer: keep each skill folder in ~/.agents/skills. Codex and OpenCode already read that folder. Claude Code doesn't, so add a link to each skill in ~/.claude/skills. Keep the frontmatter to name and description, and give each skill the same name as its folder. Checked on November 8, 2026 against Claude Code 2.1.319, Codex 0.167.0 and OpenCode 2.2.6.
Where Claude Code, Codex and OpenCode look for skills
| Folder | Claude Code | Codex | OpenCode |
|---|---|---|---|
~/.agents/skills | No | Yes | Yes |
.agents/skills in a project | No | Yes | Yes |
~/.claude/skills | Yes | No | Yes |
.claude/skills in a project | Yes | No | Yes |
~/.codex/skills | No | Yes, but deprecated | No |
~/.config/opencode/skills, .opencode/skills | No | No | Yes |
| How you call a skill | /skill-name | $skill-name or /skills | @skill-name |
All three also pick a skill on their own when your request matches its description. For project folders, each tool checks the folder you start it in and every parent folder up to the root of the repository.
- Claude Code is the one that won't find
.agents/skills. Its skills docs list only.claudefolders, and its memory docs say it doesn't read "anything under a.agents/directory". The request to change this, issue #31005, is still open with more than 450 upvotes. - Codex moved to
.agents/skills. Its docs list.agents/skillsin the project,~/.agents/skillsfor you and/etc/codex/skillsfor admins. The old~/.codex/skillsfolder is marked deprecated in the source code but still loads. Codex doesn't read any.claudefolder. - OpenCode reads everything. Its docs list its own folders plus
.claude/skillsand.agents/skills, at both levels. This is true of the 2.2 line and of the 1.19 line, which is still being released.
Set up one shared folder
Use ~/.agents/skills as the one real copy. Codex and OpenCode read it without any setup, and npx skills and the Codex importer both write there. Claude Code then gets a link to each skill.
- Move your skills into
~/.agents/skills. For each skill you wrote in~/.claude/skillsor~/.codex/skills, move its folder, for examplemv ~/.claude/skills/release-notes ~/.agents/skills/. Leave~/.claude/skills/syncedwhere it is: Claude Code keeps the skills from your claude.ai account there. - Link each skill into Claude Code's folder. On macOS or Linux, run this. It skips any skill that already has a folder or link there.
mkdir -p ~/.agents/skills ~/.claude/skills
for d in ~/.agents/skills/*/; do
n=$(basename "$d")
[ -e ~/.claude/skills/"$n" ] || ln -s ~/.agents/skills/"$n" ~/.claude/skills/"$n"
done- On Windows, use junctions. They don't need admin rights. In PowerShell:
New-Item -ItemType Directory -Force "$HOME\.claude\skills" | Out-Null
Get-ChildItem "$HOME\.agents\skills" -Directory | ForEach-Object {
$link = "$HOME\.claude\skills\$($_.Name)"
if (-not (Test-Path $link)) { New-Item -ItemType Junction -Path $link -Target $_.FullName | Out-Null }
}- Check each tool. Start a new session. In Claude Code, type
/and look for the skill. In Codex, run/skills. In OpenCode, type@.
For a project the team shares, do the same inside the repository: keep the skills in .agents/skills and link each one from .claude/skills with a relative link such as ln -s ../../.agents/skills/release-notes .claude/skills/release-notes.
A few things to get right:
- Link each skill, not the whole folder. If
~/.claude/skillsitself points to~/.agents/skills, Claude Code'ssyncedand.trashfolders land in the shared folder, and Codex and OpenCode may load the synced skills too. The cost is that you run the loop again after you add a skill. - Link folders, not SKILL.md files. Codex follows a linked skill folder, but it ignores a SKILL.md that is itself a link. OpenAI closed that request as not planned.
- Don't commit links for Windows teammates. Git for Windows checks links out as small text files by default. Create the links with a setup script, or commit real copies.
- OpenCode sees each skill twice and keeps one. It reads both the link in
~/.claude/skillsand the folder in~/.agents/skills. Version 2.2 merges them quietly. Version 1.19 logs a duplicate warning; setOPENCODE_DISABLE_CLAUDE_CODE_SKILLS=1to stop OpenCode reading the Claude folders.
Write a skill all three can load
This SKILL.md loads the same way in Claude Code, Codex and OpenCode. It lives in ~/.agents/skills/release-notes/SKILL.md, with a references and a scripts folder next to it.
---
name: release-notes
description: "Drafts release notes from git commits since the last tag. Use when the user asks for a changelog, release notes, or what changed since the last release."
license: MIT
---
# Release notes
1. Find the last tag: git describe --tags --abbrev=0
2. List the commits since it: git log <tag>..HEAD --oneline --no-merges
3. Group the commits into Added, Changed and Fixed, using references/format.md in this skill's folder.
4. To save a file, run scripts/write_notes.sh from this skill's folder.- Give the skill the same name as its folder. Use lowercase letters, numbers and hyphens, up to 64 characters. Codex goes by
name, OpenCode 2.2 goes by the folder name, and Claude Code accepts either. If the two differ, you call the same skill by different names in each tool. Codex skips a skill whose name is longer than 64 characters. - Keep the description under 1,024 characters and put the words people will use first. That is the standard's limit, and Codex cuts longer descriptions to that length in its skill list.
- Put any value that contains a colon followed by a space in quotes. Otherwise the frontmatter may not parse. Codex repairs most of these cases, Claude Code loads the skill as if it had no frontmatter, and OpenCode can skip the skill without an error.
- Make scripts find their own folder. The tools run commands from your project folder, not the skill's folder. A script that starts with
cd "$(dirname "$0")"works from anywhere. - Leave "claude" out of the name if you might upload the skill to claude.ai or the Claude API. Both reject names that contain "claude" or "anthropic".
What doesn't carry over from Claude Code
Codex reads only name, description and a short description under metadata. OpenCode 2.2 also reads metadata and disable-model-invocation. Both ignore every other field without an error, so a Claude Code skill loads but behaves differently.
| In a Claude Code skill | Claude Code | Codex | OpenCode 2.2 |
|---|---|---|---|
allowed-tools | Works | Ignored | Ignored |
disable-model-invocation: true | Works | Ignored; use agents/openai.yaml | Works |
context: fork, agent, model, hooks, argument-hint | Works | Ignored | Ignored |
$ARGUMENTS and $1 in the text | Replaced with your arguments | Left as text | Left as text |
!`command` in the text | Runs before Claude reads the skill | Left as text | Left as text |
${CLAUDE_SKILL_DIR} | Replaced with the skill's folder | Left as text | Left as text |
- Write steps that work without these features. Say "run scripts/check.sh from this skill's folder" instead of using
${CLAUDE_SKILL_DIR}, and ask for missing details in plain words instead of using$ARGUMENTS. Claude Code follows the same instructions fine. - To stop Codex from calling a skill on its own, add
agents/openai.yamlto the skill folder withallow_implicit_invocation: falseunderpolicy. You can still call it with$skill-name. allowed-toolsdoesn't limit what Codex or OpenCode can do. Thenpx skillscompatibility table lists it as supported in both, but neither tool's skill loader reads the field. Set limits in each tool's own permission settings.
Tools that install skills for several agents
| Tool | How it shares skills | What to watch |
|---|---|---|
npx skills (Vercel, 38,500 stars) | Copies a skill into .agents/skills and links each agent's folder to it. Uses junctions on Windows. --copy makes separate copies. | Same layout as this guide. Run npx skills add owner/repo, with -g for your user folder. |
| skillshare (2,900 stars) | Keeps one source folder and links it into each tool's folder; junctions on Windows. | Updated November 6, 2026. Installs with Homebrew or a shell script. |
Codex /import | Copies skills from .claude/skills into .agents/skills. | The copies drift from the originals. It also changes "claude" to "Codex" in the text, including inside file paths (open issue #40147). |
| openskills (10,800 stars) | Copies skills into .claude/skills and lists them in AGENTS.md. | No release since January 2026. Its --universal option writes to .agent/skills, which none of the three tools read. |
What about CLAUDE.md and AGENTS.md?
Codex and OpenCode read AGENTS.md, and OpenCode falls back to CLAUDE.md when a project has no AGENTS.md. Since version 2.1.277, Claude Code reads AGENTS.md too, but only in a project that has no CLAUDE.md or CLAUDE.local.md. If you keep both files, open /config in Claude Code and set Project instructions to load both. Put rules that every agent should follow in AGENTS.md, and keep CLAUDE.md for anything only Claude Code needs.
Use the same skills from Orkas
Orkas is a desktop app that runs Claude Code, Codex, OpenCode, OpenClaw and Hermes as agents in one chat. It uses the copies installed on your computer and your own sign-ins. Here is how your skills reach those agents.
- The shared folder keeps working. Each CLI still reads its own skill folders when Orkas starts it, so the setup above works unchanged. Orkas adds its own MCP server next to yours instead of replacing them.
- Skills you add in Orkas reach every CLI. Every CLI that Orkas starts gets an MCP server named
orkas. It lists your Orkas skills, loads one when the task matches, and runs the scripts the skill comes with. Ask Claude Code to use a skill by name, or let it find the right one. - Orkas reads your Claude Code and Codex skills too. Skills in
~/.claude/skillsand~/.codex/skillsappear under Skills ▸ Global folders, and the Commander can use them in its own tasks. Orkas only reads these folders and never changes them. It doesn't read~/.agents/skillsdirectly, so with the setup above it finds your skills through the links in~/.claude/skills.
To add a skill pack from GitHub, paste the repository link into a chat with the Commander and ask it to install the pack. If the pack needs to install dependencies, Orkas asks you first. The skills appear under Skills ▸ External packages. From then on, Claude Code, Codex and OpenCode can all use them when you run them from Orkas, without copying anything into three folders. To start, choose Connect agents in the sidebar and pick the CLIs you have installed. For a full walkthrough of running two of them side by side, see How to Run Claude Code and Codex Together.
Which setup should you choose?
| Your situation | Do this |
|---|---|
| You use two or three of these tools on one computer | Keep skills in ~/.agents/skills and link each into ~/.claude/skills |
| A team shares skills in a repository | Commit .agents/skills; create the .claude/skills links with a setup script |
| You install skills from GitHub often | npx skills add, which sets up the same links for you |
| You only use Claude Code and OpenCode | Keep using ~/.claude/skills; OpenCode already reads it |
| You run the CLIs from Orkas | Install skill packs once in Orkas; every CLI it starts can use them |
Questions
Can I use Claude skills in Codex?
Yes. Codex reads the same SKILL.md format, but it looks in ~/.agents/skills and .agents/skills, not in ~/.claude/skills. Move the skill folders there and link them back for Claude Code, or run /import in Codex to copy them. Fields such as allowed-tools and features such as $ARGUMENTS only work in Claude Code.
Does OpenCode read Claude Code skills?
Yes. OpenCode reads ~/.claude/skills and .claude/skills as well as ~/.agents/skills and its own folders, so Claude Code skills appear without any setup. To turn that off, set OPENCODE_DISABLE_CLAUDE_CODE_SKILLS=1.
Does Claude Code read .agents/skills?
No. As of November 2026, Claude Code reads skills only from .claude folders, and its docs say it ignores everything under a .agents directory. The request to add it, issue #31005, is still open. Until then, link each skill from ~/.claude/skills.
Why doesn't Codex see my linked skill?
Codex follows a linked skill folder, but it ignores a SKILL.md file that is itself a link. Link the whole folder instead. A relative link is resolved from the folder the link sits in, so check that the path is right from there.
Why does a skill show up twice?
Usually because two copies are in folders the tool reads. Codex still reads the old ~/.codex/skills folder as well as ~/.agents/skills, and lists both copies of a skill with the same name. Keep one copy in ~/.agents/skills and delete the other.
To compare the three tools beyond skills, see Claude Code vs Codex vs OpenCode.