Skip to content
Short answer: In the MCP vs Skills question, use an MCP server when Claude needs to reach a live tool or data source, such as GitHub, a database or a browser. Use a Skill when Claude already has the access it needs and just needs to know how you want the work done. Most real setups use both: the server gives Claude hands, the Skill gives it a playbook.
The two are easy to mix up because both “extend Claude”. They solve different problems. An MCP server is a program that exposes tools and data to an AI app. A Skill is a folder of instructions, and sometimes scripts, that Claude reads when a task calls for it. If MCP is new to you, start with our guide on what MCP servers are, then come back. You can browse real examples in the Klyvoo MCP servers directory.
Skills vs MCP servers at a glance
| Agent Skill | MCP server | |
|---|---|---|
| What it is | A folder with a SKILL.md file of instructions | A program that exposes tools, resources and prompts |
| Main job | Teach Claude a workflow or way of working | Connect Claude to an outside system |
| Where it runs | Claude reads it from the filesystem | Locally (stdio) or remotely (Streamable HTTP) |
| Always in context | Only its name and description | Depends on the app; tool definitions can add up |
| Loaded when | The task matches the description | The server is connected |
| Needs sign-in | No, unless its scripts call a service | Often, for the service behind it |
| Open standard | Documented by Anthropic | Yes, an open protocol |
| Main risk | A bad Skill can steer Claude into harmful actions | A server can expose you to prompt injection |
What is an MCP server?
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. The MCP site compares it to a USB-C port for AI apps: one standard plug, many devices. An MCP server is the program on the other end of that plug.
A server can offer three kinds of things. Tools are functions the AI can run, such as “create an issue” or “query this table”. Resources are data the AI can read, such as file contents or database records. Prompts are reusable templates. Most servers you will meet are mostly tools.
Servers run in two ways. A local server is a program on your own machine that talks to the app over stdio. A remote server runs on the vendor’s side and talks over Streamable HTTP. For the commands to connect either kind, see our guide to adding MCP servers to Claude Code or the one for Claude on the web and Desktop.
What is a Claude Skill?
Anthropic describes Skills as reusable, filesystem-based resources that give Claude domain-specific expertise: workflows, context and best practices. Each Skill is a directory with a SKILL.md file. That file starts with a short block of metadata, a name and a description, followed by the instructions in plain Markdown.
The Skill can also bundle other files: reference documents, templates and scripts. Claude opens them only when the instructions point to them.
In Claude Code, a Skill is just a folder you create. Anthropic’s docs list two places for it:
| Location | Path | Who gets it |
|---|---|---|
| Personal | ~/.claude/skills/<skill-name>/SKILL.md | All your projects on this machine |
| Project | .claude/skills/<skill-name>/SKILL.md | Everyone who uses the repository, once you commit it |
A minimal Skill looks like this:
---
name: release-notes
description: Writes release notes from merged pull requests. Use when the user asks for a changelog or release notes.
---
## Instructions
List merged changes by area. Keep each line under 15 words. End with a Breaking changes section.
You can run a Skill yourself with /skill-name, or let Claude pick it when your request matches the description. The name can be up to 64 characters of lowercase letters, numbers and hyphens. The description can run to 1,024 characters. That description matters most, because it is how Claude decides whether the Skill applies.
How each one loads into Claude’s context
This is the biggest practical difference. A model has a limited context window, so what sits in it all day has a cost.
Skills use what Anthropic calls progressive disclosure, in three levels:
- Metadata. Only the
nameanddescriptionare always loaded, about 100 tokens per Skill. - Instructions. The body of
SKILL.mdloads when the Skill is triggered, and Anthropic puts it under 5k tokens. - Resources and scripts. Bundled files load only when needed. A script runs through bash and only its output enters the context, not its code.
Anthropic’s own summary is that you can install many Skills without a context penalty, because until one triggers only its name and description occupy space.
MCP works differently. When a client connects to a server, it asks the server which tools it has. The AI app then combines them into one list the model can use. A server with a large toolset puts a lot of definitions in that list. Our GitHub MCP server page notes that the full toolset adds many tools to the AI’s context. Some clients reduce this by loading tools on demand, and the MCP docs describe progressive tool discovery for that purpose. Claude Code also caps what a tool can send back: it warns when MCP tool output goes past 10,000 tokens and limits it to 25,000 by default.
Skills are cheap to keep around. Servers cost more, but they do something a text file cannot.
When to use an MCP server
Pick a server when the job needs something outside the chat:
- Live data. Reading issues, tickets, pages or database rows that change daily.
- Actions in another app. Creating a pull request, posting a message or updating a record.
- A real tool. Driving a browser with Playwright MCP or reading your files with Filesystem MCP.
- One setup shared across apps. The same server works in any client that speaks MCP, so you build or install it once.
A Skill cannot fetch your Notion pages or open a pull request on its own. Its scripts could, in principle, but then you are running your own code and handling the credentials yourself, which is what a server already does.
When to use a Skill
Pick a Skill when Claude already has the access and needs better direction:
- A repeatable workflow. Your release checklist, your code review rules or your report format.
- House style. Voice, templates and naming rules you would otherwise paste into every chat.
- Deterministic steps. A bundled script that validates a file the same way every time.
- Specialist know-how. Reference notes on your database schema or API that Claude should consult only when relevant.
Because a Skill is plain text in a folder, it is easy to read, edit and review in version control. That is much harder with a running server.
Where each one works
Skills are not tied to Claude Code. Anthropic’s docs list three surfaces, and a Skill does not carry over between them:
| Surface | How you add Skills | Notes |
|---|---|---|
| Claude Code | Folders in ~/.claude/skills/ or .claude/skills/ | Full network access, like any program on your computer |
| claude.ai | Upload a zip in Settings > Features | Pro, Max, Team and Enterprise plans with code execution on; individual to each user |
| Claude API | Upload through the Skills API | Needs the code execution tool; no network access; shared across the workspace |
MCP is wider. It is an open protocol, and the MCP site lists Claude, ChatGPT, Visual Studio Code and Cursor among the apps that support it. Our MCP vs API post covers how it differs from calling a web API yourself.
Using Skills and MCP servers together
The two fit together well. Here is a pattern you can copy.
Say you review pull requests every week. Connect the GitHub server so Claude can read the pull request, its diff and its comments. Then write a Skill called pr-review that tells Claude what to check, what tone to use and how to format the summary. The server supplies the facts. The Skill supplies the standard.
The same split works for a browser. Playwright MCP lets Claude open and click through a page. A Skill can hold your test conventions: which pages matter, what a pass looks like and how to report a failure. If you use Claude Code for this kind of work, our best AI coding agents page shows how it compares with the others.
A good rule is to ask one question per need. “Does Claude need to reach something?” Add a server. “Does Claude need to know how we do it?” Write a Skill.
Security: treat both like installed software
Both add power, so both need trust.
Anthropic tells you to use Skills only from sources you trust, meaning ones you wrote or got from Anthropic. A malicious Skill can direct Claude to call tools or run code in ways that do not match its stated purpose, so Anthropic’s advice is to treat a Skill like installing software. Read every file in it, including scripts.
For servers, the Claude Code docs say to verify that you trust each server before connecting it, and warn that servers that fetch external content can expose you to prompt injection. The same care applies to an unfamiliar Skill that fetches data from a URL.
What neither one does
- Neither guarantees good judgment. A Skill is guidance and a server is access. Claude still decides how to use them, so check important results.
- A Skill does not give new access. If the task needs data Claude cannot reach, instructions will not fix that.
- A server does not teach your process. It can hand Claude a tool, but not your team’s way of using it.
- Neither syncs everywhere. A Skill you upload to claude.ai is not available in Claude Code, and the reverse is also true.
FAQ
Are Claude Skills a replacement for MCP servers?
No. They do different jobs. A Skill gives Claude instructions and optional scripts, while an MCP server connects Claude to an outside tool or data source. Many setups use both.
What is the difference between MCP and Skills?
MCP is an open protocol for connecting AI apps to external systems through tools, resources and prompts. A Skill is a folder with a SKILL.md file that Claude reads when a task matches its description. One is about access, the other is about method.
Which uses more context, a Skill or an MCP server?
Usually the server. A Skill keeps only its name and description loaded, about 100 tokens, until it triggers. A connected server adds its tool definitions to the list the model sees, and a big toolset adds many of them.
Can I use Skills in Claude Code and on claude.ai?
Yes, but you set them up separately. Claude Code reads Skills from folders on your machine, while claude.ai needs a zip uploaded in Settings. Anthropic says custom Skills do not sync across surfaces.
Do I need to code to make a Skill?
No. A Skill can be a single Markdown file with a name, a description and instructions. Code only comes in if you want to bundle scripts.
Are Skills and MCP servers safe?
Only as safe as their source. Anthropic recommends using Skills only from sources you trust and treating them like installed software. For servers, the Claude Code docs advise trusting each server before you connect it, since content it fetches can carry prompt injection.
Want a server to pair with your first Skill? Browse the Klyvoo MCP servers directory and filter by what you need Claude to reach, then compare setup steps in our Claude Code MCP guide.