XGSE: Extreme General Skill Engine

Overview

• The General Skill Engine is Implementation of Anthropic-Agent-Skills Protocol
• To empower your AI agents with a modular skill framework that supports dynamic skill discovery, progressive context loading, planning, and task execution.
• The Agent Skills Framework implements a multi-level progressive context loading mechanism that efficiently manages skill discovery and execution:

1. Level 1 (Metadata): Load all skill names and descriptions
2. Level 2 (Retrieval): Retrieve and load SKILL.md when relevant with the query
3. Level 3 (Resources): Load additional files (references, scripts, resources) only when referenced in SKILL.md
4. Level 4 (Analysis|Planning|Execution): Analyze the loaded skill context, plan the execution steps, and run the necessary scripts

This approach minimizes resource consumption while providing comprehensive skill capabilities.

XGSE Features

• 📜 Standard Skill Protocol: Fully compatible with the Anthropic Skills protocol
• 🧠 Heuristic Context Loading: Loads only necessary context—such as References, Resources, and Scripts on demand
• 🤖 Autonomous Execution: Agents autonomously analyze, plan, and decide which scripts and resources to execute based on skill definitions
• 🔍 Skill Management: Supports batch loading of skills and can automatically retrieve and discover relevant skills based on user input
• 🛡️ Code Execution Environment: Optional local direct code execution or secure sandboxed (E2B...) execution, with automatic dependency installation and environment isolation
• 📁 Multi-file Type Support: Supports documentation, scripts, and resource files
• 🧩 Extensible Design: The skill data structure is modularized, with implementations such as SkillSchema and SkillContext provided for easy extension and customization

Run Skill Prerequisites

• If Execution code on localhost with uv command, must install 'pip' in .venv:

  source ".venv/bin/activate"
  python -m ensurepip --upgrade 

• Use E2B Sandbox, must config 'E2B_API_KEY' in env
• Use LiteLLM, must config 'LLM_API_KEY' in env
• Can use OpenSkill download Claude skills to local

Examples

• skill: Run skill on local host or sandbox