The short answer
Choose a Skill when what's difficult is the method, meaning the steps, the format and the checks. Choose MCP when what's difficult is reaching a system such as Jira, an API or a database. Most real workflows end up with one of each.
MCP and Skills compared
In brief: anybody can write a Skill, which is a folder of guidance, while an MCP server is software that must run somewhere, normally with credentials.
| Agent Skill | MCP server | |
|---|---|---|
| Made of | SKILL.md, scripts and reference files in a folder | A server offering prompts, resources and tools |
| Runs on | Guidance inside the agent; scripts on your computer | A process on your machine (stdio) or a hosted service (HTTP) |
| When it loads | At first just its name and description; the body once needed | Most clients read every tool definition at the start |
| What the agent gains | Judgement: steps, format, a standard to meet | Reach: actions and data that live elsewhere |
| Who builds it | Anybody who writes Markdown | Mostly developers, often with credentials |
MCP, a Skill, or both: which do you need?
Work out what the agent lacks. If it has the material but handles the task differently each time, a Skill is missing. If it cannot see the data, MCP is missing.
Just a Skill
The agent has the material; only the method is absent. Say, turning notes you paste in into a brand deck.
Just MCP
Live data once, with no set process. “What happened on ticket 482?”
Both together
A report you rerun each week on fresh figures.
Do Claude Skills work with MCP? Yes, like this
They do. A Skill can name the connected MCP tools to use, the order to use them in, and how to verify the results. It does not open the connection itself, so the MCP server must already be set up. Here is a weekly report Skill relying on two servers:
## Steps
1. Call the Linear tools and collect every issue closed in the last seven days.
2. Call the analytics tools and fetch this week's active users.
3. Write the results into references/weekly-report.md.
4. Mark each metric that changed by more than 10%.Codex lets a Skill declare the MCP servers it relies on in agents/openai.yaml; a Codex plugin may also bundle MCP connections with Skills.
Can Skills be delivered by an MCP server?
Yes. Through MCP's official Skills extension (io.modelcontextprotocol/skills), a server can offer Skills written to the Agent Skills spec. The extension is finished, but agent support is still arriving, so most people add Skills as folders for now.
Which uses up more context?
An unused Skill costs very little. An unused MCP server usually costs far more.
~1 sentence
per installed Skill, until it is called on
Every tool
description that each connected MCP server exposes, loaded at the start by most clients
So connect only the MCP servers a task really calls for, not all of them.
Common questions
If I have Skills, do I still need MCP?
Only when the agent must get into a system it has no other way to reach. A Skill cannot grant access; it explains how to make good use of access you already have.
Do Claude Skills work with MCP?
Yes. A Skill's instructions can point Claude at specific connected MCP tools and say how to handle the output.
Can Skills come from an MCP server?
Yes, by way of the official MCP Skills extension. Support across agents is still being added.
Does MCP still matter now that Skills are here?
It does. The two handle separate problems: a Skill holds the method, and MCP holds the connection.
Sources
Did this guide help?
