Which AI Coding Agents Are Compatible with k-skill?

k-skill supports Claude Code, Codex, OpenCode, OpenClaw/ClawHub, Hermes Agent, Vercel Agent Skills, and any AI coding agent that respects the ~/.claude/skills or ~/.agents/skills directory conventions.

k-skill is a universal skill library designed to function across multiple AI coding agent environments. According to the NomaDamas/k-skill source code, the repository explicitly defines agent compatibility in AGENTS.md and automatically adapts its installation layout based on whether the target agent follows Claude-style or agents-compatible conventions.

Supported AI Coding Agents

The k-skill repository distinguishes between two primary agent families based on their expected skill installation paths. As documented in AGENTS.md, this bifurcation ensures seamless integration without manual configuration.

Claude-Style Agents

Claude-style agents expect skills to reside in ~/.claude/skills/<skill-name>/ and utilize top-level SKILL.md files for discovery. The following agents follow this convention:

• Claude Code – The primary target environment with native ~/.claude/skills support.
• Codex – Shares the Claude Code layout and functions identically within the ~/.claude/skills directory.
• OpenCode – Treated as a Claude-compatible environment using the same installation path.
• OpenClaw / ClawHub – Recognized as Claude-compatible agents per the README.md, which states these agents are supported ("Claude Code, Codex, OpenCode, OpenClaw/ClawHub 등 각종 코딩 에이전트 지원합니다").

Agents-Compatible Agents

Agents-compatible agents utilize the ~/.agents/skills/<skill-name>/ directory structure. k-skill supports:

• Hermes Agent – Uses the ~/.agents/skills home install path with SKILL.md discovery.
• Vercel Agent Skills – Implements the same agents-compatible layout as Hermes Agent.
• Other custom agents – Any agent respecting the ~/.agents/skills convention can automatically discover k-skill packages.

Installation Paths and Detection Logic

k-skill's detection mechanism relies on standard filesystem conventions rather than proprietary APIs. According to AGENTS.md and docs/install.md, the system differentiates agent types through their home directory layouts.

The detection logic operates as follows:

1. Claude-style detection checks for the presence of ~/.claude/ and installs to ~/.claude/skills/<skill-name>/.
2. Agents-compatible detection falls back to ~/.agents/skills/<skill-name>/ for broader agent support.

Each skill package includes a skill.json metadata file (located at the skill root, e.g., zipcode-search/skill.json) that instructs agents how to expose the skill. The packages/k-skill-cli/README.md documents the CLI wrapper that abstracts these path differences.

Cross-Agent Usage Examples

k-skill provides both CLI wrappers and direct script execution methods that work identically across all supported agents.

Command Line Execution

Execute any k-skill capability using the npx wrapper without worrying about agent-specific paths:

# Example: Search for Korean scholarships

npx -y @nomadamas/k-skill@0 exec korean-scholarship-search \
    scripts/korean_scholarship_search.py -- query "2024 장학금"

Agent-Specific Installation

For persistent integration, install skills to the agent-specific directory using the provided scripts:

Claude Code installation:

~/.claude/skills/korean-privacy-terms/scripts/install.sh

Hermes Agent installation:

~/.agents/skills/korean-privacy-terms/scripts/install.sh

Programmatic Python Integration

Access k-skill capabilities from Python scripts using subprocess calls that remain agent-agnostic:

import subprocess
import json

def run_skill(skill, script, *args):
    cmd = [
        "npx", "-y", "@nomadamas/k-skill@0", "exec",
        skill, f"scripts/{script}", "--"
    ] + list(args)
    result = subprocess.run(cmd, capture_output=True, text=True)
    return json.loads(result.stdout)

# Example: Get nearby cheap gas stations

data = run_skill("cheap-gas-nearby", "cheap_gas_nearby.py", "--location", "Seoul")
print(data)

Summary

• k-skill supports Claude Code, Codex, OpenCode, OpenClaw/ClawHub, Hermes Agent, and Vercel Agent Skills.
• Claude-style agents use ~/.claude/skills/<skill-name>/ for skill installation.
• Agents-compatible agents use ~/.agents/skills/<skill-name>/ for broader compatibility.
• The AGENTS.md file defines detection logic, while skill.json files provide per-skill metadata.
• Cross-agent usage is possible via the @nomadamas/k-skill CLI wrapper or direct script execution.

Frequently Asked Questions

What is the difference between Claude-style and agents-compatible installation?

Claude-style installation targets agents like Claude Code and Codex that expect skills in ~/.claude/skills/<skill-name>/, while agents-compatible installation uses ~/.agents/skills/<skill-name>/ for agents like Hermes and Vercel. Both methods rely on top-level SKILL.md discovery, but the path differs based on the agent's home directory conventions as defined in AGENTS.md.

Can I use k-skill with custom or self-hosted agents?

Yes. Any AI coding agent that respects the ~/.agents/skills/ directory layout and SKILL.md discovery mechanism is automatically compatible. The repository's detection logic in AGENTS.md treats any agent following these filesystem conventions as a valid target, requiring no proprietary API integration.

How does k-skill handle skill metadata across different agents?

Each skill package contains a skill.json file (e.g., zipcode-search/skill.json) that exposes metadata about the skill's capabilities, entry points, and requirements. Agents read this file to understand how to execute the skill, making the metadata format agent-agnostic while the installation paths remain agent-specific.

Where can I find detailed installation instructions for my specific agent?

Detailed step-by-step installation guides for each supported agent type are documented in docs/install.md within the repository. This file provides specific commands for Claude Code, Hermes Agent, and other compatible environments, ensuring proper setup regardless of which AI coding agent you use.

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