Which AI Coding Agents Are Compatible with Claude Skills?

Nine major AI coding agents currently support Claude Skills, including Claude Code, Claude .ai, the Claude API, OpenAI Codex, Cursor, Gemini CLI, Antigravity, Windsurf, and OpenCode, using a portable folder format containing a SKILL.md file.

Claude Skills provide an open-standard specification for packaging reusable LLM behaviors into portable, version-controlled folders. According to the ComposioHQ/awesome-claude-skills repository, this format enables cross-platform compatibility because any agent capable of scanning a directory for SKILL.md files and exposing the instructions to an LLM can execute the defined workflows without platform-specific modifications.

Anthropic Ecosystem Support

The Claude Skills format originated within Anthropic's ecosystem and enjoys first-class support across their entire product line.

Claude Code

As detailed in README.md at line 349, Claude Code loads skills from the ~/.config/claude-code/skills/ directory. When you start the CLI, the agent scans this location and activates skills automatically when your prompt matches the skill's description metadata.

mkdir -p ~/.config/claude-code/skills/
cp -r my-skill ~/.config/claude-code/skills/
claude   # start Claude Code – the skill appears automatically when relevant

Claude .ai Web Interface

The web interface, documented at line 41 of the repository's README.md, displays skills with a puzzle-piece icon (🧩) in the UI. Users can add skills from the marketplace or upload custom folders directly through the interface, making the functionality accessible without terminal access.

Claude API

For programmatic access, line 71 of README.md specifies that you pass a list of skill IDs via the skills parameter in client.messages.create. This allows you to invoke skills within automated pipelines or custom applications.

import anthropic

client = anthropic.Anthropic(api_key="YOUR_KEY")
resp = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    skills=["my-skill-id"],               # skill ID from the marketplace or your repo

    messages=[{"role":"user","content":"Help me generate a changelog"}],
)
print(resp)

Third-Party AI Coding Agents

Because the Skills format is intentionally open and tool-agnostic, multiple third-party agents have implemented support, as listed at line 36 of the repository's README.md.

OpenAI Codex

OpenAI Codex accepts the same skill folder structure, importing the SKILL.md file and invoking the described workflow when relevant. The JavaScript SDK provides a direct method for skill execution:

const codex = require("@openai/codex");
await codex.runSkill({
  skillPath: "./my-skill",
  prompt: "Create a README for a new npm package",
});

Cursor

Cursor, the AI-native code editor, reads skill folders and provides autocomplete-style prompts based on the skill's description. You integrate skills via the CLI:

cursor skill add ./my-skill
cursor chat "Write a unit test for function X"

Gemini CLI

Google's Gemini CLI executes skills by feeding the SKILL.md contents directly into Gemini's prompting pipeline:

gemini skill import ./my-skill
gemini chat "Summarize the changelog from recent git commits"

Antigravity

The Antigravity agent loads skill metadata via its plugin system and triggers execution when conversation context matches the skill's activation criteria:

from antigravity import load_skill
load_skill("./my-skill")

# later in a conversation

agent.execute("Generate a project scaffold")

Windsurf

Windsurf treats skills as first-class plugins invocable from the command line:

windsurf install ./my-skill
windsurf run "Create a Dockerfile for a Node.js app"

OpenCode

Formerly known as OpenAI Codex CLI, OpenCode (referenced at line 190 of README.md) maintains the same compatibility, reading skill folders and executing described steps using the identical runtime logic as its predecessor.

The Unified Runtime Model

All compatible agents share a common four-stage execution model, as outlined in the repository documentation:

1. Discovery – The agent scans a directory (or marketplace) for folders containing a SKILL.md file.
2. Metadata loading – Only the YAML front-matter (name, description) is loaded initially, requiring approximately 100 tokens.
3. Lazy loading – The full markdown body and auxiliary scripts are fetched only when the agent determines the skill is relevant to the current context.
4. Execution – The agent follows the step-by-step instructions inside SKILL.md, calling any declared tools such as MCP servers, APIs, or local scripts as needed.

This architecture ensures that skills remain lightweight until needed and execute consistently regardless of which agent loads them.

Core File Structure

For a skill to be compatible across all these agents, it must adhere to the minimal contract defined in the repository's key files:

• SKILL.md – The core specification containing YAML front-matter and markdown instructions (see examples like connect-apps/SKILL.md at line 126).
• README.md – Repository overview including the authoritative list of compatible agents (line 1).
• CONTRIBUTING.md – Guidelines ensuring new skills maintain cross-agent compatibility.
• connect-apps-plugin/.claude-plugin/plugin.json – Example plugin configuration demonstrating how skills integrate with Claude Code and similar agents.

Summary

• Nine agents currently support Claude Skills: Claude Code, Claude .ai, Claude API, OpenAI Codex, Cursor, Gemini CLI, Antigravity, Windsurf, and OpenCode.
• The format requires only a SKILL.md file with YAML front-matter in a named folder, making it truly portable.
• Lazy loading ensures skills consume minimal tokens until explicitly invoked.
• All agents follow the same discovery → metadata loading → execution pipeline, ensuring consistent behavior across platforms.

Frequently Asked Questions

What makes Claude Skills compatible across different AI coding agents?

Claude Skills use a simple folder-based format with a standardized SKILL.md file containing YAML front-matter and markdown instructions. Because the specification is open and tool-agnostic, any agent that implements directory scanning for SKILL.md files and can expose the content to an LLM can execute the skill, regardless of the underlying model provider.

Do I need to modify skills to work with different agents?

No. According to the ComposioHQ/awesome-claude-skills repository, skills follow a universal contract that requires no platform-specific modifications. A single skill folder works identically in Claude Code, OpenAI Codex, Cursor, and other compatible agents because all implementations respect the same YAML front-matter structure and execution instructions.

How do I distribute Claude Skills to my team?

You can distribute skills via version-controlled repositories, shared network drives, or private marketplaces. For Claude Code specifically, place skills in ~/.config/claude-code/skills/. For other agents like Cursor or Windsurf, use their respective CLI commands (cursor skill add or windsurf install) pointing to the skill folder path.

Are there performance limits for skill folders?

The repository specifies that only the YAML front-matter (approximately 100 tokens) loads initially during the discovery phase. The full markdown body and any auxiliary scripts remain unloaded until the agent determines the skill is relevant, keeping memory and token usage minimal until execution is actually required.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}