openskills MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

Rust runtime — fast, predictable
TS and Python bindings for embedding into any agent
Progressive disclosure: metadata → body → scripts only when needed
Sandboxing via Seatbelt (macOS) and experimental WASM
Cross-agent: not tied to Claude Code

라이브 데모

실제 사용 모습

openskills-skill.replay ▶ 준비됨

0/0

설치

클라이언트 선택

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "openskills-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Geeksfino/openskills",
        "~/.claude/skills/openskills"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop → Settings → Developer → Edit Config 열기. 저장 후 앱 재시작.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "openskills-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Geeksfino/openskills",
        "~/.claude/skills/openskills"
      ],
      "_inferred": true
    }
  }
}

Cursor는 Claude Desktop과 동일한 mcpServers 스키마 사용. 프로젝트 설정이 전역보다 우선.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "openskills-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Geeksfino/openskills",
        "~/.claude/skills/openskills"
      ],
      "_inferred": true
    }
  }
}

Cline 사이드바의 MCP Servers 아이콘 클릭 후 "Edit Configuration" 선택.

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "openskills-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Geeksfino/openskills",
        "~/.claude/skills/openskills"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "openskills-skill",
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Geeksfino/openskills",
        "~/.claude/skills/openskills"
      ]
    }
  ]
}

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

{
  "context_servers": {
    "openskills-skill": {
      "command": {
        "path": "git",
        "args": [
          "clone",
          "https://github.com/Geeksfino/openskills",
          "~/.claude/skills/openskills"
        ]
      }
    }
  }
}

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add openskills-skill -- git clone https://github.com/Geeksfino/openskills ~/.claude/skills/openskills

한 줄 명령. claude mcp list로 확인, claude mcp remove로 제거.

사용 사례

실전 활용법: openskills

How to embed Agent Skills into your own agent runtime

👤 Engineers building custom LLM agents ⏱ ~120 min advanced

언제 쓸까: You're writing an agent from scratch and want a skill system that already understands SKILL.md.

사전 조건

Rust toolchain (for building) or prebuilt binding — rustup install stable; or use the TS/Python binding
Repo cloned — git clone https://github.com/Geeksfino/openskills ~/.claude/skills/openskills

흐름

Install the binding

Install the openskills Python or TS binding and point it at my skills directory.✓ 복사됨

→ Runtime loaded; skills enumerated
Expose to your agent

Wire the runtime into my agent's planner so it considers skills at each step.✓ 복사됨

→ Agent has a SkillRegistry abstraction
Run sandboxed

Enable Seatbelt sandboxing for skill scripts on macOS.✓ 복사됨

→ Skill scripts run under sandbox profile

결과: Your custom agent can load and run SKILL.md bundles like Claude Code does.

함정

Loading all scripts at startup kills memory — Rely on progressive disclosure — metadata first
Sandbox denies legitimate operations — Audit the sandbox profile; narrow-then-widen rather than default-deny everything

함께 쓰기: filesystem

Run a skill in isolation to test it outside Claude Code

👤 Skill authors who want cross-agent validation ⏱ ~45 min intermediate

언제 쓸까: You want to prove your skill works without the specific harness of Claude Code.

흐름

Load the skill

Load ~/.claude/skills/my-skill/ with openskills and print its metadata.✓ 복사됨

→ Parsed frontmatter + body
Invoke with a prompt

Feed it a sample task and observe what would get loaded.✓ 복사됨

→ Disclosure trace: what the agent would see

결과: Confidence your skill behaves portably.

함정

Skill depends on Claude Code internals — Refactor the skill to rely only on the public SKILL.md interface

조합

다른 MCP와 조합해 10배 효율

openskills-skill + skill-optimizer-skill

Validate skills portably before optimizing them

Load the skill with openskills, then hand off to skill-optimizer for patch suggestions.✓ 복사됨

openskills-skill + filesystem

Operate on the skills directory directly

List all SKILL.md in ~/.claude/skills/, summarize each.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출	비용
skill-loader	skills dir	Agent startup	minimal
progressive-disclosure	skill id + stage	During agent reasoning	minimal
sandbox-runner	script + policy	Any time a skill spawns a subprocess	script runtime

비용 및 제한

운영 비용

API 쿼터: none
호출당 토큰: runtime is prompt-light; you control what's disclosed
금액: free, open source
팁: Keep progressive disclosure on — loading every script upfront blows context

보안

권한, 시크릿, 파급범위

자격 증명 저장: none at runtime level

데이터 외부 송신: none beyond what scripts themselves do

WASM sandboxing is experimental — don't rely on it for adversarial scripts
Seatbelt profiles should be tightened per your use case

문제 해결

자주 발생하는 오류와 해결

Binding import fails

Ensure the native binary matches your platform; rebuild from source if needed.

Sandbox blocks expected syscalls

Iterate on the Seatbelt profile; log denials to find the right allow list.

대안

openskills 다른 것과 비교

대안	언제 쓰나	단점/장점
Claude Code native skill loader	You're only running inside Claude Code	No cross-agent portability

openskills