Back to Blog

Skills vs MCP Servers vs Plugins — What's the Difference?

By SkillShelf
#skills#mcp-servers#plugins#claude-code#comparison

If you’re building with AI agents, you’ve probably heard three terms thrown around: MCP servers, skills, and plugins. They all extend what an AI agent can do — but they’re not the same thing, and picking the wrong one wastes time.

Here’s the short version: MCP servers connect agents to external data. Skills give agents specialized instructions. Plugins are dead.

Now the longer version.

What Are MCP Servers?

MCP stands for Model Context Protocol. An MCP server is a lightweight process that exposes external tools and data sources to AI agents through a standardized protocol. Think of it as a USB port for your AI — plug in a server, and the agent gains new capabilities.

An MCP server for GitHub lets your agent read repos, create issues, and open pull requests. A server for PostgreSQL lets it query your database. A server for DataForSEO lets it pull keyword research data.

The key characteristics of MCP servers:

  • They connect to external systems. Databases, APIs, file systems, SaaS tools — anything with an interface.
  • They’re language-agnostic. MCP servers can be written in TypeScript, Python, Rust, or anything that speaks the protocol.
  • They’re reusable across clients. The same MCP server works with Claude Code, Claude Desktop, Cursor, Windsurf, and any other MCP-compatible client.
  • They run as separate processes. The server runs alongside your AI client, communicating over stdio or HTTP.

MCP servers don’t tell the agent how to think about the data — they just make the data available. The agent decides what to do with it.

Browse MCP servers on MCP Atlas or explore the MCP Servers category here on SkillShelf.

What Are Claude Code Skills?

A Claude Code skill is a markdown file that gives Claude specialized instructions for a specific task. It lives in your project’s .claude/skills/ directory and activates when you invoke it with a slash command or when Claude automatically matches the current task to the skill’s description.

Skills don’t connect to external systems. They shape how the agent thinks and works. A /blog-post skill knows your site’s front matter format, SEO requirements, and cross-linking rules. A /twitter skill knows your brand voice and thread structure. A /marketing skill holds your entire content strategy.

The key characteristics of skills:

  • They’re prompt files. A skill is a markdown document with instructions, examples, and guardrails.
  • They’re project-scoped. Skills live in your repo and carry your project’s context.
  • They have no runtime dependencies. No npm packages, no API keys, no processes to manage. Just a file.
  • They’re Claude Code-specific. Skills work in Claude Code. They don’t transfer to Cursor, Windsurf, or other clients.

Skills are opinionated by design. An MCP server gives you access to keyword data. A skill tells the agent how to use that data to write a blog post that passes your SEO checklist.

Explore Claude Code skills in the Skills category on SkillShelf.

What Were Plugins?

Plugins were OpenAI’s first attempt at extending ChatGPT, launched in March 2023. They let developers build integrations that ChatGPT could call during a conversation — booking flights, searching the web, running code.

Plugins are gone. OpenAI shut them down in April 2024, replacing them with Custom GPTs and GPT Actions.

Why did they fail?

  • Walled garden. Plugins only worked inside ChatGPT. Building one meant committing to a single platform with no portability.
  • Discovery was broken. The plugin store was hard to navigate and most plugins had negligible usage.
  • No standard protocol. Each plugin had its own API spec. There was no shared protocol that other AI clients could adopt.
  • Limited developer control. You couldn’t customize how the AI used your plugin beyond the API description.

MCP learned from these mistakes. The Model Context Protocol is open, client-agnostic, and doesn’t depend on any single AI provider. That’s why MCP servers are thriving where plugins died.

Skills vs MCP Servers: Key Differences

AspectClaude Code SkillsMCP ServersPlugins (deprecated)
What they doShape agent behavior and reasoningConnect agents to external data and toolsConnected ChatGPT to third-party APIs
FormatMarkdown prompt filesStandalone server processesAPI endpoints with OpenAPI specs
RuntimeNo runtime — loaded as contextRuns as a separate process (stdio or HTTP)Hosted by the developer
PortabilityClaude Code onlyAny MCP-compatible clientChatGPT only
SetupDrop a .md file in .claude/skills/Configure in .mcp.json, install dependenciesRegister in the plugin store
DependenciesNoneMay need API keys, npm packages, DockerFull API hosting required
What they accessProject context, conversation historyExternal APIs, databases, file systemsExternal APIs
CustomizationFully customizable — it’s your promptLimited to what the server exposesLimited to API description
MaintenanceEdit the markdown fileUpdate the server packageMaintain hosted API

The simplest way to think about it: MCP servers are the data layer. Skills are the intelligence layer. Plugins were both, poorly.

When to Use Skills vs MCP Servers

Use this decision framework:

Use an MCP server when:

  • You need data from an external system (database, API, SaaS tool)
  • Multiple team members or clients need the same integration
  • The capability should work across different AI clients, not just Claude Code
  • You need real-time data that changes frequently

Use a skill when:

  • You want to standardize how the agent performs a task
  • You need project-specific knowledge baked into the workflow
  • You want guardrails, checklists, or output formats enforced
  • The task is about reasoning and structure, not data access

Use both together when:

  • You need external data and structured reasoning. For example: an MCP server pulls keyword data from DataForSEO, and a /blog-post skill uses that data to write an SEO-optimized post with the right front matter, heading structure, and cross-links.

The most powerful AI workflows combine MCP servers for data access with skills for domain expertise. Neither one replaces the other.

FAQ

What is the difference between MCP and a plugin? MCP is an open protocol that works across multiple AI clients. Plugins were proprietary to ChatGPT and are now deprecated. MCP servers are reusable, portable, and don’t require a central store or approval process.

Are Claude Code skills the same as GPTs? No. GPTs are OpenAI’s customizable ChatGPT configurations. Claude Code skills are markdown files that give Claude specialized instructions within your development workflow. Skills are version-controlled, project-scoped, and have zero platform lock-in beyond Claude Code itself.

Can I use MCP servers without Claude Code? Yes. MCP servers work with any MCP-compatible client — Claude Desktop, Cursor, Windsurf, and others. The protocol is client-agnostic by design.

Do I need to choose between skills and MCP servers? No. They serve different purposes and work best together. Use MCP servers for data, skills for workflow structure.

Where can I find MCP servers and skills? Browse MCP servers at MCP Atlas. Explore Claude Code skills and other AI agent tools here on SkillShelf.