Skip to content
Wiki

MCP connects agents to live systems; Agent Skills teach them procedures

DrFritzi · Reviewed · Updated 28 Sept 2026 · Markdown

Answer

They solve different problems, so they are not rivals. MCP (Model Context Protocol) lets an agent call tools and read resources on a running server, which suits live data that changes between calls. An Agent Skill is a folder of instructions that teaches an agent a procedure, which suits checklists, brand guides and repeatable workflows. When a procedure needs live data, write a skill that tells the agent how to use a specific MCP server. Checked 2026-09-28 against MCP specification 2026-07-28 and the Agent Skills site.

Details

What each one is

A skill is a folder that contains a SKILL.md file. The file has YAML frontmatter with name and description (both required), followed by Markdown instructions. The folder may also hold scripts/, references/ and assets/.

An MCP server exposes tools that a language model can invoke. The client sends tools/list to discover them and tools/call to run one, and the server does the work. Nothing in a skill runs on a server unless the skill's instructions tell the agent to call one.

The Agent Skills site says the format was originally developed by Anthropic, released as an open standard, and adopted by a growing number of agent products. Its overview page lists supported clients, and that list changes, so check it instead of trusting a copy.

How each one loads into context

Skills load in three stages, called progressive disclosure. First, at startup, the agent reads only each skill's name and description. Second, when a task matches, it reads the full SKILL.md. Third, it runs bundled scripts or opens referenced files only as needed. The specification recommends about 100 tokens for the metadata and under 5000 tokens for the body.

MCP tool definitions are different: a client typically puts the tool list in front of the model. See mcp-too-many-tools-context-bloat for that cost.

Decision table

The agent needs... Use Why
Live data that changes between calls (tickets, prices, database rows) MCP A skill file is static text; a server returns fresh results
An action in an external system (create, update, send) MCP The server performs it and returns a typed result
A procedure, review checklist or house style Skill It is instructions the agent reads, not something to call
Reusable scripts and templates Skill They ship inside the folder
A procedure that depends on live data Both A skill says when and how to call a particular MCP server

Worked example: a skill that drives an MCP server

Goal: an agent should answer questions from this wiki and always cite the page. The wiki's public MCP endpoint at https://wiki.drfritzi.at/mcp has two tools, search_answers and get_answer (checked 2026-09-28).

wiki-lookup/
└── SKILL.md
---
name: wiki-lookup
description: Answers technical questions from the public wiki through its MCP server. Use when the user asks about MCP, agents or dev tools and wants a sourced answer.
---

1. Call `search_answers` with 2-4 keywords from the question.
2. Pick the best match by title and summary. Call `get_answer` with its slug.
3. Answer from the returned text only. End with the page URL.
4. If nothing matches, say so. Do not answer from memory.

The skill holds the procedure and the citation rule. The server holds the content. The agent's client must also have the server configured, which a skill does not do for you.

Common mistakes

  • Putting volatile data in a skill. It goes stale and the agent will not notice.
  • Building an MCP server just to serve a static checklist. A skill is simpler.
  • Writing a vague skill description. The agent decides whether to activate a skill from the description alone.
  • Assuming a skill can call a server on its own. The client must connect the server first.
  • Naming a skill with uppercase letters or spaces. The specification allows lowercase letters, digits and single hyphens, and the name must match the folder name.

See also

Sources