MCP vs skills: how MCP servers and agent skills differ

An MCP server connects an AI agent to something outside it, such as an app, a database or a file system, and gives it tools to read and act there. A skill is a folder of instructions, optionally with scripts, that teaches the agent how to do a task. In short, MCP provides access and skills provide procedure, and the two work well together.

Anthropic puts the rule of thumb this way: if you are explaining how to do something, write a skill; if Claude needs to reach something, use MCP.

The difference at a glance

What an MCP server does

An MCP server exposes three kinds of building blocks to an AI application. Tools are functions the model can call, such as "create issue" or "search flights". Resources are data the application can read, such as a file or a database schema. Prompts are reusable templates a user can pick. The protocol lets any compatible client, from Claude to Cursor or VS Code, discover and use these in the same way.

The server handles the connection and the login. That makes MCP the right choice when the agent must see live data or change something in another system.

What a skill does

A skill carries knowledge rather than a connection. At the start of a session the agent sees only each skill's name and a short description. When a request matches, it reads the instructions, and it opens bundled files or runs bundled scripts only when a step calls for them.

Skills suit work that needs judgement plus a consistent method: how your team formats reports, which checks to run before a release, or how to fill in a particular form.

How they compare on the points that matter

Context cost

Both are designed to stay light. Anthropic estimates a skill's always-visible metadata at about 100 tokens. For MCP, the Claude Code documentation says tool names load at the start of a session, while each tool's full schema is fetched only when needed. Other clients may load tools differently, so check your client's documentation if you connect many servers.

Setup and maintenance

An MCP server is software with versions, dependencies and often credentials. A remote server is maintained by its provider; a local one is a program you install and update. A skill is mostly text, so it is easy to read, edit and keep in version control.

Portability

MCP is supported by many AI clients. Skills follow the open Agent Skills format, and agents such as Claude Code, Codex and Gemini CLI document support for it. Product-specific front matter fields may not carry across.

Risk

Both can lead an agent to change things. The MCP specification treats tools as arbitrary code execution and expects hosts to ask for user consent before calling them. Anthropic warns that a malicious skill can steer Claude into running code or calling tools in ways that do not match its stated purpose. The risk depends on what each one can reach, not on the format.

When to choose which

Choose an MCP server when:

• the agent needs data it cannot otherwise see, such as tickets, records or messages;
• the agent must take an action in another system;
• you are copying information into the chat by hand from another tool.

Choose a skill when:

• you keep pasting the same instructions into conversations;
• the task follows a fixed method or output format;
• you want a repeatable script run as one step of a larger job.

Using both together

The two combine naturally. The Claude Code documentation gives the example of a database server that runs queries alongside a skill that documents your schema and query patterns. Anthropic's blog describes Notion publishing a meeting-preparation skill that tells Claude which pages to search and how to lay out the result, while the Notion MCP server does the searching and saving.

Some practical patterns:

• Project tracker: an MCP server for Linear or Jira reads and creates issues; a skill holds your team's rules for writing them.
• Code hosting: a GitHub server opens pull requests; a skill sets the code review checklist.
• Design handoff: a Figma server fetches the design; a skill encodes your frontend design conventions.

When a skill refers to MCP tools, Anthropic advises writing the server name and tool name together, for example GitHub:create_issue, so the agent picks the right tool when several servers are connected.

The MCP project also lists a working group on delivering skills through MCP itself. Treat that as an area to watch rather than something to rely on today.

Safety note

Before installing either, check who publishes it, what it can change and what code it runs. Our MCP and skill security questions walk through each point.

Browse both on Kitbay

Kitbay lists MCP servers and agent skills side by side for each tool, so you can see whether a tool has a server, a skill or both. Vendor-published listings are collected on the official listings page.