Claude Skills vs MCP: What Each One Is For

Claude skills and MCP solve different problems. A skill is a folder of instructions that teaches Claude how to do a task. MCP, the Model Context Protocol, is a standard for connecting Claude to outside tools and data through servers. Use a skill when Claude needs to know a procedure, and an MCP server when Claude needs access to a system. Many setups use both.

What a skill is

A skill is a folder with a SKILL.md file: a name, a description, and written instructions, optionally with scripts and reference files next to it. Claude keeps the name and description in context, and loads the full instructions when a task matches or when you call the skill by name.

Nothing runs in the background. A skill is text on disk until it is used. For a fuller explanation, see what are Claude skills.

What MCP is

The Model Context Protocol is an open-source standard for connecting AI applications to tools. An MCP server is a program that exposes tools, and Claude Code's documentation describes the result plainly: MCP servers give Claude access to your tools, databases and APIs.

With servers connected, Claude can do things such as read an issue from your tracker, query a database, or create a draft in your email, because the server performs the call. Servers can be remote, reached over HTTP, or local processes on your machine. In Claude Code you add one with a command:

claude mcp add --transport http notion https://mcp.notion.com/mcp

An MCP server defined by a plugin runs alongside each session where that plugin is enabled.

The difference in one table

	Skill	MCP server
What it is	A folder of instructions and files	A program that exposes tools
What it gives the agent	Knowledge of how to do something	Access to a system or data source
Made of	Markdown, plus optional scripts	Code that implements the protocol
Defined by	The Agent Skills open standard	The Model Context Protocol
Loads	Name and description at startup; body when used	Tools become callable once the server connects
Needs a running process	No	Yes, local or remote
Typical example	"How we write release notes"	"Read and write issues in our tracker"

A simple test: if you could solve the problem by handing a new colleague a written procedure, it is a skill. If you would have to give them a login to a system, it is an MCP server.

When to use a skill

Anthropic's guidance is to create a skill when you keep pasting the same instructions, checklist or multi-step procedure into chat. Good candidates:

• A review checklist your team applies to every pull request.
• The format and tone of a recurring document.
• A deployment procedure with steps in a fixed order.
• Conventions for a codebase that are too long to keep in always-loaded project instructions.

Skills can include scripts, so they are not limited to prose. A skill can ship a script that validates a file or fills in a form, and the agent runs it instead of writing the code again each time.

When to use an MCP server

The Claude Code documentation suggests connecting a server when you find yourself copying data into chat from another tool, such as an issue tracker or a monitoring dashboard. Good candidates:

• Reading and updating tickets.
• Querying a database.
• Pulling designs, documents or messages from a hosted service.
• Anything that needs authentication to a third-party system.

The server handles the connection and authentication. The agent only sees tools it can call.

Using them together

The two fit together because they cover different halves of a job. The MCP server makes an action possible, and the skill says when and how to take it.

For example, a server might expose tools for your issue tracker. A skill can then describe your triage process: which fields to read, how to label an issue, when to ask a person. Without the skill, Claude can call the tools but has to guess your process. Without the server, Claude knows the process but cannot reach the tracker.

In Claude Code, a plugin can package both. A plugin is a directory that can contain skills, subagents, hooks and MCP servers, installed as one unit, so a skill and the server it relies on arrive together.

Cost in context

Both features are built to avoid filling the context window before they are needed.

For skills, only the name and description of each one are loaded at startup, and the body is read on activation.

For MCP in Claude Code, tool search is on by default. It defers tool definitions until Claude needs them, so only tool names and server instructions load when a session starts.

The practical difference is elsewhere. A skill, once invoked, stays in the conversation as text. An MCP tool call returns data, which can be large. Neither is free once it is in use.

Trust and security

The risks differ in kind.

A skill is instructions and files. A malicious one can direct the agent to run commands or use tools in ways that do not match its stated purpose. Anthropic's advice is to audit a skill from an unknown source before using it: read SKILL.md and every bundled script.

An MCP server is running code, and it may fetch content from elsewhere. The Claude Code documentation warns that servers which fetch external content can expose you to prompt injection, and says to verify that you trust each server before connecting it.

In both cases, install from sources you trust and read what you are installing.

Which should you reach for

• You repeat the same instructions: write or install a skill.
• You copy data from another tool into chat: connect an MCP server.
• You need both the access and the procedure: use both, or a plugin that bundles them.

To find skills, browse the directory. To compare skills with the other ways of extending Claude Code, read Claude skills vs agents, plugins and commands.