What Are Claude Skills and Plugins? A Developer’s Guide to Modular AI Workflows

Claude Skills are modular, self-contained instruction packages that teach a Claude agent how to perform tasks, while Claude Plugins are MCP gateways that let the agent execute real-world actions across 500+ SaaS applications.

Claude Skills and Plugins form the extensibility layer of the ComposioHQ/awesome-claude-skills repository, separating agent workflows from live tool integrations. They let developers teach Anthropic’s Claude what to do and how to interact with external services without bloating the context window. This guide explains their architecture, loading behavior, and how to use them together.

What Are Claude Skills?

Claude Skills are modular, self-contained instruction packages that teach a Claude agent how to perform a particular class of tasks. Each skill lives in its own folder and must contain a SKILL.md file with YAML front-matter—defining name and description—followed by Markdown instructions. The repository’s skill-creator/SKILL.md file documents how optional sub-folders such as scripts/, references/, and assets/ hold reusable code and media that are fetched on demand.

Anatomy of a Skill Folder

A valid skill directory follows a simple convention:

• SKILL.md – Required. Contains YAML front-matter (name, description) and the instructional body.
• scripts/ – Optional. Holds executable code loaded lazily or run without entering context.
• references/ – Optional. Stores supplementary documentation.
• assets/ – Optional. Stores images, PDFs, or other files pulled on demand.

This folder structure keeps the repository organized and ensures the agent only materializes resources when necessary.

Progressive-Disclosure Loading Layers

According to skill-creator/SKILL.md, skills use a three-layer loading strategy to conserve the agent’s context window:

1. Metadata (name + description) – Always in context, costing roughly 100 tokens.
2. SKILL.md body – Loaded only when the agent judges the skill relevant, typically staying under 5 k tokens.
3. Bundled resources – Loaded lazily as needed; scripts can even be executed without being read into the token context.

This progressive-disclosure design is what allows the ecosystem to offer unlimited functionality while keeping the active context small.

What Are Claude Plugins?

Claude Plugins are MCP (Model Context Protocol) gateways that expose real-world actions to the agent. Rather than describing a workflow, a plugin supplies the tool layer: authenticated API calls to external services. The ComposioHQ/awesome-claude-skills repository ships a Connect Apps plugin, documented in connect-apps-plugin/README.md, that wires Claude to 500+ SaaS applications through the Composio platform.

How the Connect Apps Plugin Works

When the plugin is installed, Claude receives an API key and initiates OAuth authorization flows for each target service. After authorisation, the agent can invoke those services as normal tools. This means actions like sending emails, creating GitHub issues, or updating CRM records happen through standardised MCP tool calls rather than custom scripts.

How Skills and Plugins Work Together

Skills and Plugins are complementary. A skill describes the workflow: the steps to take, the guardrails to apply, and the format of the output. A plugin provides the execution layer: the live connection to Salesforce, Slack, Gmail, or any other supported SaaS. As outlined in the repository’s README.md, this pairing enables autonomous domains such as document processing, code generation, data analysis, and enterprise automation.

Running a Claude Skill in Python

You can load a skill by passing its identifier to the skills parameter in the Anthropic Messages API:

import anthropic

client = anthropic.Anthropic(api_key="YOUR_API_KEY")
response = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    skills=["skill-id-here"],          # replace with the skill’s identifier

    messages=[{"role": "user", "content": "Generate a summary of the attached PDF"}],
)
print(response)

This snippet demonstrates how the client pulls in the named skill’s instructions before processing the user prompt.

Installing the Connect Apps Plugin

To add real-world tool access, install the plugin from the repository root and run the interactive setup:

# Install the plugin

claude --plugin-dir ./connect-apps-plugin

# Run the interactive setup (you’ll be asked for a Composio API key)

/connect-apps:setup

Once setup completes, you can test the integration with a natural-language request:

Send me a test email at myemail@example.com

These steps are documented in connect-apps-plugin/README.md.

Summary

• Claude Skills are folder-based instruction modules built around a SKILL.md file. They use progressive-disclosure loading to keep context usage low while offering unlimited functionality.
• Claude Plugins are MCP gateways that connect the agent to live SaaS tools; the bundled Connect Apps plugin supports 500+ applications via Composio.
• Skills define workflows and guardrails; plugins provide the authenticated tool layer.
• Key reference files include skill-creator/SKILL.md for skill authors, connect-apps-plugin/README.md for plugin installation, and mcp-builder/SKILL.md for building custom MCP servers.

Frequently Asked Questions

What is the difference between a Claude Skill and a Plugin?

A Claude Skill is a modular set of instructions stored in SKILL.md that tells the agent what workflow to execute and what rules to follow. A Claude Plugin is an MCP gateway that exposes external APIs—such as email, GitHub, or Slack—as invocable tools. Skills govern behavior; plugins provide the mechanism to act.

How do I create a custom Claude Skill?

Create a directory with a SKILL.md file that includes YAML front-matter for name and description, followed by Markdown instructions. You can add optional scripts/, references/, and assets/ sub-folders for lazy-loaded resources. The skill-creator/SKILL.md guide in the repository provides the full specification and progressive-disclosure principles.

What is MCP and how does it relate to Claude Plugins?

MCP stands for Model Context Protocol, the open standard that Claude Plugins use to register and call external tools. An MCP gateway translates Claude's intent into authenticated API requests, which is how the Connect Apps plugin enables access to hundreds of SaaS services without custom wrappers. Developers looking to build custom MCP servers can consult mcp-builder/SKILL.md for implementation guidance.

How many SaaS applications does the Composio plugin support?

The Connect Apps plugin, as implemented in ComposioHQ/awesome-claude-skills, connects Claude to 500+ SaaS applications through the Composio platform. It handles OAuth flows and API key management so the agent can invoke these services as standard tools.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}