Skills

Skills vs MCPs vs Plugins: comparison

Complete comparison between Claude Code Skills, MCPs, and Plugins. When to use each one, strengths and limitations.

Three mechanisms, three missions

Claude Code offers three ways to extend its capabilities: Skills, MCPs (Model Context Protocol), and Community extensions (often loosely called "plugins"). Each serves a different need, and it's by combining them that you get the most out of the ecosystem.

'Plugins' don't exist as described elsewhere

Claude Code doesn't have /plugin install commands or an official plugin marketplace. What the community calls "plugins" are actually collections of files (agents, rules, skills) that are installed manually from GitHub or via /install-github. This guide uses the term "community extension" for accuracy.

The full analogy

Think of Claude Code as a craftsperson in their workshop.

  • Skills are their recipes and procedures: they tell them how to carry out a specific task, step by step.
  • MCPs are their connected tools: a power drill plugged into the electrical network, a welding station connected to the gas line. They provide access to external resources.
  • Community extensions are shared configurations: sets of agents, rules, and skills bundled by the community for a specific use.

Comparison table

CriteriaSkillMCPCommunity extension
DefinitionMarkdown instruction fileExternal server exposing tools via JSON-RPCShared collection of agents, rules, and skills
AnalogyCooking recipeUSB port to a serviceShared configuration pack
Format.md fileServer (Node.js, Python, etc.)Folder of .md / .mdc files
InstallationCopy into .claude/commands/claude mcp add or settings.json/install-github or manual copy
ScopeProject or userProject or globalProject or global
InvocationSlash command (/project:...)Automatic (Claude picks the tool)Automatic (active agents) or slash command
IntelligenceInstructions followed by ClaudeTools called by ClaudeInstructions interpreted by Claude
ComplexityLow (just Markdown)Medium (server to maintain)Low (text files)
SharingVia git (.claude/commands/)Via shared configurationVia GitHub
ExamplesPR Review, TDD Guide, DeployGitHub, Slack, PostgreSQLeverything-claude-code, rule collections

When to use what?

The choice depends on what you're trying to accomplish. Here's a quick decision guide.

Use a Skill when...

  • You want to standardize a process (review, deploy, onboarding)
  • You want to teach a specific behavior to Claude
  • You want something simple to create and maintain
  • The workflow is self-contained (no need to call an external service)
  • You want to share a process with the team via git
# Perfect examples for a Skill
/project:review-pr          # Standardized review process
/project:create-component   # Component creation per your conventions
/user:tdd-guide             # Personal TDD workflow
/project:deploy-checklist   # Deployment checklist

Use an MCP when...

  • You need to access an external service (API, database, SaaS)
  • You want Claude to read from or write to a third-party system
  • The tool needs to be always available (not manually invoked)
  • You need a standardized communication protocol
# Perfect examples for an MCP
claude mcp add github       # Access repos, PRs, GitHub issues
claude mcp add slack        # Read and send Slack messages
claude mcp add postgres     # Query a database
claude mcp add context7     # Access up-to-date documentation

Dive deeper into MCPs

Check our complete MCP guide to understand how they work in detail, and the top recommended MCPs to discover the most useful ones.

Use a community extension when...

  • You want to benefit from a complete pack (agents + skills + templates) maintained by the community
  • You don't want to create your files from scratch and prefer starting from a proven base
  • You want to share a configuration with your team or the community
  • The configuration is complex and requires multiple coordinated files
# Install from GitHub via /install-github
/install-github punkpeye/everything-claude-code   # Complete agent and skill pack


# Or manually
git clone https://github.com/punkpeye/everything-claude-code
cp -r everything-claude-code/agents/ .claude/agents/
cp -r everything-claude-code/rules/ .claude/rules/

Dive deeper into community extensions

Check our complete guide on community extensions to understand how they work, and the top essential extensions to get started.

Winning combinations

The real power of Claude Code shows when you combine these three mechanisms. Here are the most effective combinations.

Combo 1: Full-stack development

# MCP for external context
claude mcp add github          # Access to issues and PRs
claude mcp add context7        # Up-to-date library documentation


# Community extension for agents
/install-github punkpeye/everything-claude-code


# Skills for team workflows
/project:create-api            # Create an endpoint per your conventions
/project:review-pr             # Standardized review

Result: Claude accesses GitHub issues (MCP), uses the community collection's specialized agents, and follows your team conventions (Skills).

Combo 2: Security workflow

# Security agent from Everything Claude Code
/install-github punkpeye/everything-claude-code


# Skill for the complete audit
/project:security-audit


# GitHub MCP to create tracking issues
claude mcp add github

Result: The security-reviewer agent analyzes the code, the Skill structures the audit according to your checklist, and the GitHub MCP automatically creates issues for found vulnerabilities.

Combo 3: Developer onboarding

# MCP to access documentation
claude mcp add context7


# Skill to guide the newcomer
/project:onboarding


# Community extension for specialized agents
/install-github punkpeye/everything-claude-code

Result: The MCP provides access to up-to-date documentation (Context7), the Skill guides the new developer through the onboarding steps, and specialized agents assist them on specific tasks.

Start simple, enrich progressively

No need to install everything at once. Start with one or two Skills for your daily workflows, add MCPs for the services you use, then explore plugins when you're comfortable. The ecosystem is designed for progressive adoption.

Decision tree

Not sure which mechanism to use? Follow this guide:

Do you need to access an external service?

Yes -> Use an MCP. It's the only mechanism that connects Claude Code to third-party services (GitHub, Slack, databases, APIs).

No -> Move to the next question.

Do you want to standardize a process or workflow?

Yes -> Use a Skill. It's the simplest way to codify a process. A Markdown file is all you need.

No -> Move to the next question.

Do you need a complete pack of agents and rules?

Yes -> Use a community extension (like everything-claude-code). These collections bundle agents, skills, and templates into a coherent folder, maintained by the community. Install them with /install-github or manually.

No -> A simple Skill is probably enough. Start there and evolve if needed.

FAQ

Can I use all three at the same time?

Absolutely. The three mechanisms are designed to be complementary. A community extension can include Skills, and those Skills can be used alongside MCP tools. Combining all three gives the best results.

Can a Skill call an MCP?

Not directly. A Skill gives instructions to Claude, and Claude can then decide to use an MCP tool if it's relevant. The Skill doesn't "know about" MCPs, but Claude intelligently makes the connection.

Does a community extension replace Skills and MCPs?

No. A community extension often includes Skills (Markdown files in a commands/ folder), but it doesn't replace MCPs. MCPs are still needed for connecting to external services. Think of a community extension as a set of configuration files that enriches Claude Code without requiring an external server.

How do I choose between a project Skill and a personal Skill?

  • Project Skill (.claude/commands/): for everything involving conventions and workflows shared by the team. Versioned with the code.
  • Personal Skill (~/.claude/commands/): for your individual work preferences. Not shared.

Do Skills consume context?

Yes, when a Skill is invoked, its content is injected into Claude's context. That's why it's important to keep Skills concise and focused. A 100-200 line Skill is ideal. Beyond 500 lines, consider splitting it.

Can you create a custom MCP?

Yes, an MCP is a server that implements the JSON-RPC protocol. You can create one with the MCP SDK in TypeScript, Python, or other languages. Check our MCP guide for more details.

What's the performance impact?

  • Skills: minimal impact (just text injected into context)
  • MCPs: depends on the external service (network latency, API response time)
  • Community extensions: variable impact (added agents and rules consume context every session)

Going further

Previous

Creating custom Skills