Library

The Library

A meta-skill for private-first distribution of agentics (skills, agents, prompts, tools, and MCP servers) across agents, devices, and teams.

Variables

Update these after forking and cloning the library repo.

• LIBRARY_REPO_URL: https://github.com/<your-org>/<your-library-repo>.git
• LIBRARY_YAML_PATH: ~/.claude/skills/skill-library/library.yaml
• LIBRARY_SKILL_DIR: ~/.claude/skills/skill-library/

How It Works

The Library is a catalog of references to your agentics. The library.yaml file points to where skills, agents, prompts, and tools live (local filesystem or GitHub repos). Nothing is fetched until you ask for it.

The library.yaml is a catalog, not a manifest. Entries define what's available — not what gets installed. You pull specific items on demand with /library use <name>.

Commands

Cookbook

Each command has a detailed step-by-step guide. Read the relevant cookbook file before executing a command.

When a user invokes a /library command, read the matching cookbook file first, then execute the steps.

Source Format

The source field in library.yaml supports these formats (auto-detected):

• /absolute/path/to/SKILL.md — local filesystem
• https://github.com/org/repo/blob/main/path/to/SKILL.md — GitHub browser URL
• https://raw.githubusercontent.com/org/repo/main/path/to/SKILL.md — GitHub raw URL

Both GitHub URL formats are supported. Parse org, repo, branch, and file path from the URL structure. For private repos, use SSH or GITHUB_TOKEN for auth automatically.

Important: The source points to a specific file (SKILL.md, AGENT.md, TOOL.md, or prompt file). We always pull the entire parent directory, not just the file.

Source Parsing Rules

Local paths start with / or ~:

• Use the path directly. Copy the parent directory of the referenced file.

GitHub browser URLs match https://github.com/<org>/<repo>/blob/<branch>/<path>:

• Parse: org, repo, branch, file_path
• Clone URL: https://github.com/<org>/<repo>.git
• File location within repo: <path>

GitHub raw URLs match https://raw.githubusercontent.com/<org>/<repo>/<branch>/<path>:

• Parse: org, repo, branch, file_path
• Clone URL: https://github.com/<org>/<repo>.git
• File location within repo: <path>

Generic Git hosting URLs (Gitea, GitLab 등) match https://<host>/<org>/<repo>/blob/<branch>/<path> where <host> is NOT github.com:

• Parse: host, org, repo, branch, file_path
• Clone URL: https://<host>/<org>/<repo>.git
• File location within repo: <path>

GitHub Workflow

When working with GitHub sources, prefer gh api for accessing single files (e.g., reading a SKILL.md to check metadata). For pulling entire skill directories, clone into a temp dir per the steps below.

Fetching (use):

1. Clone the repo with git clone --depth 1 <clone_url> into a temporary directory
2. Navigate to the parent directory of the referenced file
3. Copy that entire directory to the target local directory
4. The temporary directory is cleaned up automatically

Pushing (push):

1. Clone the repo with git clone --depth 1 <clone_url> into a temporary directory
2. Overwrite the skill directory in the clone with the local version
3. Stage only the relevant changes: git add <skill_directory_path>
4. Commit with message: library: updated <skill name> <what changed>
5. Push to remote
6. The temporary directory is cleaned up automatically

Typed Dependencies

The requires field uses typed references to avoid ambiguity:

• skill:name — references a skill in the library catalog
• agent:name — references an agent in the library catalog
• prompt:name — references a prompt in the library catalog
• tool:name — references a CLI tool in the library catalog
• mcp:name — references an MCP server in the library catalog

When resolving dependencies: look up each reference in library.yaml, fetch all dependencies first (recursively), then fetch the requested item.

Target Directories

By default, items are installed to the default directory from library.yaml:

default_dirs:
    skills:
        - default: .claude/skills/
        - global: ~/.claude/skills/
        - codex_global: ~/.codex/skills/
        - gemini_global: ~/.gemini/skills/
        - gsd_global: ~/.gsd/agent/skills/
    agents:
        - default: .claude/agents/
        - global: ~/.claude/agents/
    prompts:
        - default: .claude/commands/
        - global: ~/.claude/commands/
    tools:
        - default: .claude/tool-cli/
        - global: ~/.claude/tool-cli/

• If the user says "global" or "globally", use the global directory.
• If the user specifies a custom path, use that path.
• Otherwise, use the default directory.

Scope Field

각 스킬 엔트리에 scope 필드를 지정할 수 있다:

• scope: project (기본값) — /library use 시 .claude/skills/에만 설치
• scope: global — /library use시 모든 에이전트 타겟(claude, codex, gemini, gsd)에 설치. /library sync-all의 대상이 된다.

scope 필드가 없으면 project로 간주한다.

Library Repo Sync

The library skill itself lives in <LIBRARY_SKILL_DIR> as a cloned git repo. When running add (which modifies library.yaml), always:

1. git pull in the library directory first to get latest
2. Make the changes
3. git add library.yaml && git commit && git push

This keeps the catalog in sync across devices.

MCP Server Management

MCP servers differ from file-based items (skills, agents, prompts, tools). They are config entries stored in ~/.claude.json under the mcpServers key, while the catalog definition lives in library.yaml under library.mcps.

Config Format

Each MCP entry in library.yaml has:

- name: server-name
  description: What the MCP does
  type: stdio|http
  config:
    # For stdio: command, args, env
    # For http: url, headers

Security: Secret Placeholders

Sensitive values (API keys, tokens) are stored as {{PLACEHOLDER}} in the catalog. When installing via mcp-use, the user is prompted for actual values.

Auto-detect fields: any config key containing key, token, secret, password, credential, or auth (case-insensitive) is treated as sensitive.

JSON Safety

When modifying ~/.claude.json:

1. Always backup first: cp ~/.claude.json ~/.claude.json.bak
2. Use python3 for JSON round-trip (never edit JSON as text)
3. Validate after write: python3 -c "import json; json.load(open('$HOME/.claude.json'))"
4. Restore from backup if validation fails

Example Filled Library File

default_dirs:
  skills:
    - default: .claude/skills/
    - global: ~/.claude/skills/
  agents:
    - default: .claude/agents/
    - global: ~/.claude/agents/
  prompts:
    - default: .claude/prompts/
    - global: ~/.claude/prompts/
  tools:
    - default: .claude/tool-cli/
    - global: ~/.claude/tool-cli/

library:
  skills:
    - name: firecrawl
      description: Scrape, crawl, and search websites using Firecrawl CLI
      source: /Users/me/projects/tools/skills/firecrawl/SKILL.md

    - name: meta-skill
      description: Creates new Agent Skills following best practices
      source: /Users/me/projects/tools/skills/meta-skill/SKILL.md

    - name: diagram-kroki
      description: Generate diagrams via Kroki HTTP API supporting 28+ languages
      source: https://github.com/myorg/private-skills/blob/main/skills/diagram-kroki/SKILL.md
      requires: [skill:firecrawl]

    - name: green-screen-captions
      description: Generate and burn AI-powered captions onto green screen videos
      source: https://raw.githubusercontent.com/myorg/video-tools/main/skills/green-screen-captions/SKILL.md
      requires: [agent:video-processor, prompt:caption-style]

  agents:
    - name: video-processor
      description: Processes video files with ffmpeg and whisper transcription
      source: /Users/me/projects/tools/agents/video-processor/AGENT.md

    - name: code-reviewer
      description: Reviews code for quality, security, and performance
      source: https://github.com/myorg/agent-configs/blob/main/agents/code-reviewer/AGENT.md

  prompts:
    - name: caption-style
      description: Style guide for generating video captions
      source: /Users/me/projects/content/prompts/caption-style.md

    - name: commit-message
      description: Standardized commit message format for all projects
      source: https://github.com/myorg/team-prompts/blob/main/prompts/commit-message.md

  tools:
    - name: jk
      description: Jenkins CLI for job management and pipeline control
      source: /Users/me/projects/tools/tool-cli/jk/TOOL.md

    - name: browser-cli
      description: Headless browser automation for web scraping
      source: https://github.com/myorg/cli-tools/blob/main/tool-cli/browser-cli/TOOL.md

  mcps:
    - name: context7
      description: Up-to-date library documentation lookup via Context7
      type: http
      config:
        type: "http"
        url: "https://mcp.context7.com/mcp"
        headers:
          CONTEXT7_API_KEY: "{{CONTEXT7_API_KEY}}"

    - name: svgmaker
      description: SVG generation via SVGMaker API
      type: stdio
      config:
        type: "stdio"
        command: "npx"
        args: ["-y", "@genwave/svgmaker-mcp"]
        env:
          SVGMAKER_API_KEY: "{{SVGMAKER_API_KEY}}"