Zum Inhalt springen

Skills / skill fetch

skill fetch

Multi-registry skill discovery and installation for AI coding agents — search 9 sources, score, paginate, and install agent skills with security labels

22von @girofuvor 18d aktualisiertMITGitHub →

Installation

Kompatibilitaet

Claude CodeCodexGeminiCursor

Beschreibung

skill-fetch

Cross-platform skill discovery and installation for AI coding agents

Search, score, and install AI agent skills from 9 registries in parallel — works across Claude Code, Cursor, Codex, Gemini CLI, Windsurf, and Amp.

Quick Start

  1. Install — pick any method from Installation below
  2. Set up API keys — run /fetch-skill-config (interactive, recommended) or configure manually
  3. Search & install skills/fetch-skill react native animation

Features

  • 9 Search Sources — SkillsMP (semantic + keyword), GitHub, Anthropic Skills, ClawSkillHub, skills.sh, PolySkill, SkillHub, Skills Directory
  • Cross-Platform — Works on 6+ AI coding agents with automatic tool adaptation
  • Multi-Variant AI Search — 3 query variants fired in parallel, improving recall significantly
  • Quality Scoring — 0-100 composite score: Relevance (40) + Freshness (25) + Community (20) + Trust (15) + External Bonus (5)
  • Security Labels — 5 trust tiers: Official, Verified, Partial, Unverified, Security Concerns
  • 6-Category Security Scan — Destructive commands, RCE, data exfiltration, system modification, obfuscation, prompt injection
  • Prompt Injection Detection — 5 sub-categories (PI-1~PI-5): direct override, hidden role markers, encoding tricks, indirect injection, social engineering
  • Integrity Verification — SHA-256 hash recorded at install, tamper detection on future loads
  • Paginated Results — Browse 5 at a time with c to continue; install by number from any page
  • Local/Global Install — Choose project-level or user-level installation (mandatory prompt)
  • Deduplication — Same skill across registries is merged; similar descriptions are flagged
  • Bundled API Scripts — Pre-built shell scripts for SkillHub and Skills Directory APIs (no API key exposure)

Installation

Choose the method that fits your setup:

Option 1: Plugin (Best experience, Claude Code)

# Add the skill-fetch marketplace
claude plugin marketplace add girofu/skill-fetch

# Install skill-fetch plugin
claude plugin install skill-fetch@skill-fetch

Provides auto-updates, /fetch-skill command, and full plugin integration.

Option 2: npx skills add (Node.js)

npx skills add girofu/skill-fetch

Works with Claude Code, Cursor, and Codex.

Option 3: curl | sh (Universal, zero dependencies)

curl -fsSL https://raw.githubusercontent.com/girofu/skill-fetch/main/install.sh | bash

Auto-detects installed agents and installs for all of them. Specify a single agent:

curl -fsSL https://raw.githubusercontent.com/girofu/skill-fetch/main/install.sh | bash -s -- --agent claude

Option 4: Python installer

python3 -c "$(curl -fsSL https://raw.githubusercontent.com/girofu/skill-fetch/main/install.py)"

Or download and run:

curl -fsSL https://raw.githubusercontent.com/girofu/skill-fetch/main/install.py -o install.py
python3 install.py --agent cursor

Option 5: Manual (git clone)

git clone https://github.com/girofu/skill-fetch.git
# Claude Code
cp -r skill-fetch/skills/skill-fetch ~/.claude/skills/skill-fetch

# Cursor
cp -r skill-fetch/skills/skill-fetch ~/.cursor/skills/skill-fetch

# Other agents: replace ~/.cursor/ with ~/.codex/, ~/.gemini/, ~/.windsurf/, or ~/.amp/

⚙️ Setup API Keys

Important: API keys unlock the full 9-source search. Without them, only Sources 3-8 are available (still useful, but fewer results).

Interactive Setup (Recommended)

Run /fetch-skill-config in Claude Code. It will:

  • Show which keys are currently configured
  • Walk you through setting each key with links to get them
  • Automatically install the SkillsMP MCP server if needed
  • Write everything to ~/.claude/skills/.fetch-config.json

Manual Setup

🔑 SkillsMP API Key (Sources 1-2) — Recommended

SkillsMP provides the best semantic search with AI-powered matching. Get your free API key:

  1. Go to skillsmp.com and create an account
  2. Copy your API key
  3. Register the MCP server:
claude mcp add --scope user skillsmp -- npx -y skillsmp-mcp-server --env SKILLSMP_API_KEY=your_key_here
  1. Restart Claude Code — SkillsMP tools will be available in the next session

🔑 SkillHub & Skills Directory (Sources 8-9) — Optional

For additional coverage, create ~/.claude/skills/.fetch-config.json:

{
  "SKILLSMP_API_KEY": "your-skillsmp-key",
  "SKILLHUB_API_KEY": "sk-sh-your-key-here",
  "SKILLS_DIRECTORY_API_KEY": "sk_live_your-key-here"
}

Sources 3-8 work without any API keys. Source 9 (Skills Directory) requires a key.


Supported Agents

| Agent | Plugin | npx | curl/sh | Python | Manual | |-------|--------|-----|---------|--------|--------| | Claude Code | ✅ | ✅ | ✅ | ✅ | ✅ | | Cursor | — | ✅ | ✅ | ✅ | ✅ | | Codex | — | ✅ | ✅ | ✅ | ✅ | | Gemini CLI | — | — | ✅ | ✅ | ✅ | | Windsurf | — | ✅ | ✅ | ✅ | ✅ | | Amp | — | — | ✅ | ✅ | ✅ |

Usage

# Search for skills (Claude Code)
/fetch-skill react native animation

# Install from URL
/fetch-skill https://github.com/owner/repo

# Auto-triggered by skill-eval hook
# (no manual invocation needed)

In other agents, the skill activates automatically when referenced in context.

How It Works

┌─────────────────────────────────────────────────────┐
│                   User Query                         │
│              "testing React Native"                  │
└──────────────────────┬──────────────────────────────┘
                       │
        ┌──────────────┼──────────────────┐
        ▼              ▼                  ▼
   ┌─────────┐   ┌──────────┐      ┌──────────┐
   │SkillsMP │   │SkillsMP  │      │SkillsMP  │
   │AI Var A  │   │AI Var B  │      │AI Var C  │
   └────┬─────┘   └────┬─────┘      └────┬─────┘
        │              │                  │
        └──────┬───────┘──────────┬───────┘
               ▼                  ▼
   ┌───────────────┐    ┌─────────────────┐
   │ SkillsMP KW   │    │    GitHub        │
   │ + Anthropic    │    │    + PolySkill   │
   └───────┬───────┘    └────────┬────────┘
           │                     │
   ┌───────┴───────┐    ┌───────┴────────┐
   │  ClawSkillHub │    │  skills.sh     │
   └───────┬───────┘    └───────┬────────┘
           │                     │
   ┌───────┴───────┐    ┌───────┴────────┐
   │  SkillHub     │    │ Skills         │
   │  (API script) │    │ Directory      │
   └───────┬───────┘    └───────┬────────┘
           │                     │
           └──────────┬──────────┘
                      ▼
              ┌───────────────┐
              │  Deduplicate  │
              │  Score & Rank │
              │  (0-100)      │
              └───────┬───────┘
                      ▼
              ┌───────────────┐
              │  Security     │
              │  Scan (A-F)   │
              └───────┬───────┘
                      ▼
              ┌───────────────┐
              │  Display 5    │
              │  per page     │
              └───────┬───────┘
                      ▼
              ┌───────────────┐
              │  User picks   │
              │  → Install    │
              │  → Hash       │
              └───────────────┘

Cross-platform note: SkillsMP sources require the MCP server + API key (see API Key Setup). On other platforms, the flow starts from GitHub + supplementary sources.

Search Sources

| # | Source | Method | Type | Availability | |---|--------|--------|------|-------------| | 1 | SkillsMP (semantic) | skillsmp_ai_search MCP × 3 variants | Primary | Claude Code + MCP | | 2 | SkillsMP (keyword) | skillsmp_search MCP | Primary | Claude Code + MCP | | 3 | GitHub | gh search code --filename SKILL.md (primary) + gh search repos (supplementary) | Primary | All agents | | 4 | Anthropic Skills | gh search code in anthropics/skills | Official | All agents | | 5 | ClawSkillHub | npx -y clawhub search | Supplementary | Agents with npx | | 6 | skills.sh | HTTP API / WebFetch | Supplementary | All agents | | 7 | PolySkill | npx -y @polyskill/cli search (single keyword) | Supplementary | Agents with npx | | 8 | SkillHub | Bundled shell script (API) or CLI fallback | Supplementary | All agents | | 9 | Skills Directory | Bundled shell script (API) | Supplementary | Configured agents |

All sources are searched in parallel. Sources fail gracefully — if any is unavailable, the search continues with remaining sources.

API Key Setup

See ⚙️ Setup API Keys above for detailed instructions.

Quality Scoring

Each result receives a composite score (0-100):

| Dimension | Weight | How | |-----------|--------|-----| | Relevance | 0-40 | LLM judges description-to-task semantic match | | Freshness | 0-25 | Time since last GitHub push | | Community | 0-20 | Star count (log scale) | | Trust | 0-15 | Source credibility tier | | External Bonus | 0-5 | Security/quality signals from PolySkill, SkillHub, Skills Directory |

Grade labels:

| Score | Label | |-------|-------| | 85+ | 🟢 Strongly Recommended | | 70-84 | 🟢 Recommended | | 55-69 | 🟡 Worth Considering | | 40-54 | 🟡 Marginal | | <40 | 🔴 Not Recommended |

Security

Security Labels

| Label | Criteria | |-------|----------| | 🔒 Official | From anthropics/skills repo | | 🔒 Verified | SkillsMP stars ≥ 50 + securityGrade A/B + scan clean | | ⚠️ Partial | Lower stars but scan clean, or standard frontmatter present | | ⚠️ Unverified | Direct URL or no external signals | | ⚠️ Security Concerns | Scan found issues or securityGrade D/F |

6-Category Security Scan

All skills are scanned before (or immediately after) installation:

| Category | Name | Severity | |----------|------|----------| | A | Destructive Commands | Critical | | B | Remote Code Execution | Critical | | C | Data Exfiltration | High | | D | System Modification | High | | E | Obfuscation | Medium | | F | Prompt Injection (5 sub-types) | High |

Integrity Verification

After installation, SHA-256 hashes are recorded for all skill files. On future loads, hashes are compared to detect tampering.

Permissions Declaration (Advisory)

Skills may optionally declare their required permissions (network, filesystem-write, shell-commands, external-urls) in SKILL.md frontmatter. The scanner flags mismatches between declared and actual behavior.

File Structure

skill-fetch/
├── .claude-plugin/
│   └── plugin.json                   # Claude Code plugin manifest
├── commands/
│   ├── fetch-skill.md               # /fetch-skill slash command
│   └── fetch-skill-config.md        # /fetch-skill-config API key setup
├── skills/
│   └── skill-fetch/
│       ├── SKILL.md                  # Main skill (cross-platform)
│       ├── references/
│       │   ├── search-sources.md     # Source-specific commands & dedup rules
│       │   ├── quality-signals.md    # Scoring algorithm + security labels
│       │   ├── interaction-patterns.md # Output templates, security scan, prompt injection
│       │   └── platform-adapters.md  # Cross-platform tool mapping
│       └── scripts/
│           ├── fetch-skillhub.sh     # SkillHub API search (reads key from config)
│           └── fetch-skills-directory.sh # Skills Directory API search
├── install.sh                        # Universal bash installer
├── install.py                        # Python installer
├── CHANGELOG.md                      # Version history
├── README.md                         # This file
└── LICENSE                           # MIT

Requirements

Minimum (any agent):

  • Shell with curl or WebFetch for HTTP-based searches

Full experience (Claude Code):

  • SkillsMP MCP server + API key from skillsmp.com (see API Key Setup)
  • GitHub CLI (gh) for GitHub and Anthropic Skills search
  • Node.js for npx-based searches (ClawSkillHub, PolySkill, SkillHub CLI)
  • ~/.claude/skills/.fetch-config.json for SkillHub and Skills Directory APIs (optional)

License

MIT

Aehnliche Skills