MohamedHamed19m/Skill_GEMINI_Extension

Skill: GEMINI Extension

An extension that contains an MCP (Model Context Protocol) server for loading skills on demand.

Overview

This project provides a flexible and dynamic way to provide "skills" to an AI agent. This Gemini extension includes an MCP (Model Context Protocol) server for loading skills on demand and for custom commands. This helps AI agents manage their context window by loading skills as needed.

As a Gemini Extension

This project is intended to be used as a Gemini extension. The gemini-extension.json file defines how the extension is loaded and configured. It includes an MCP server and custom commands that are started automatically by the Gemini CLI.

The MCP server is defined in mcp_app/skills_server.py, and custom commands are located in the commands directory. The skills themselves are located in the skills directory.

1- Installation

This project uses uv for package and environment management.

For Developers (Local Setup)

1. Create and Sync the Virtual Environment:

   Run the following command to create a virtual environment named .venv and install the dependencies from pyproject.toml:

   uv venv
   uv sync
   

2. Activate the Virtual Environment:

   Before running the server, activate the environment.

   On Windows:

   .venv\Scripts\activate
   

3. Run the Server (for local testing/development):

   Once the environment is activated, you can start the server:

   python mcp_app/skills_server.py
   

For Users (Gemini CLI Extension Installation)

To install this repository as a Gemini CLI extension:

1. Ensure uv is installed: Follow the instructions above to install uv if you haven't already. uv is required to run the MCP server embedded within this extension.
2. Install the extension:

   gemini extensions install https://github.com/MohamedHamed19m/Skill_GEMINI_Extension
   

After installation, restart your Gemini CLI session to ensure the extension is loaded.

2. Activate

Restart the Gemini CLI. The following commands will be available:

• list_skills() - Lists all available skills with metadata (lightweight).

• search_skills(query: str, limit: int = 5) - Search for relevant skills using intelligent matching.

• load_skill(skill_name: str, force_reload: bool = False)- Load full skill content into context.

• add_skills_directory(path: str) - Add custom skills directory to scan.

Update Extension

• gemini extensions update skills

• gemini extensions update --all

Skill Structure

Skills are organized in a directory structure. The server scans the skills directory within the project.

Each skill must be in its own directory. The directory name is the skill name. Inside the skill directory, there must be a file named SKILL.md.

Example structure:

skills/
└── my-awesome-skill/
    ├── SKILL.md
    └── some-other-file.txt

In this example, the skill name is my-awesome-skill. The content of SKILL.md will be loaded as the skill's content. Any other files in the skill's directory are for reference and will not be loaded by the server.

The SKILL.md file can have a YAML frontmatter to provide metadata, for example:

---
description: This is a skill that does awesome things.
keywords: [awesome, skill]
---

The rest of the file is the skill content.

The skill name is always derived from the folder name. Any name field provided in the frontmatter will be ignored.

AI Agent Workflow

Recommended Pattern (with search)

1. Search for relevant skills:

   results = search_skills(query="How do I send ARETHIL frames?")
   # Returns: Top 3-5 relevant skills with scores

2. Load the best matches:

   load_skill(skill_name=results[0]["name"])

3. Use the content to answer the question

Fallback Pattern (without search)

1. List all skills (only if search fails):

   all_skills = list_skills()

2. Manually filter based on keywords/description

3. Load specific skills

Search Quality

• First 5-10 minutes: Uses fast keyword matching
• After model loads: Automatically upgrades to semantic search
• Transparent: AI doesn't need to know which backend is active

This project is maintained by MohamedHamed.