Skill or MCP Server: Choosing the Right Layer

---
title: "Skill or MCP Server: Choosing the Right Layer"
description: "Decide between a SKILL.md file and an MCP server for a given agent capability. The two are complementary layers, not competitors - skill catalogs (including AgentArmory) deliver skills over MCP itself. Use when the choice between layers is not obvious."
author: AgentArmory
license: Proprietary
---

# Skill or MCP Server: Choosing the Right Layer

Decide between a SKILL.md file and an MCP server for a given agent capability. Skills provide procedural knowledge as structured instructions. MCP servers provide runtime tool access. These are **complementary layers, not competitors** - a skill catalog (AgentArmory included) is itself delivered over the MCP protocol, and the most capable agent systems use both together.

This guide uses a 5-question decision framework, a decision matrix, and concrete examples to help agents (and the humans guiding them) choose the right layer for each part of a task.

## Before You Start: Layers, Not Rivals

Skills and MCP servers occupy different architectural layers. A skill encodes *what to do* (methodology, steps, decision rules). An MCP server provides *the means to do it* (tool execution, real-time data, side effects).

The confusion arises because skill catalogs - including AgentArmory - deliver skills through an MCP endpoint. A user authenticates to the catalog, queries for a skill via MCP tools, and receives skill instructions back through the same protocol. This means a single MCP connection can carry BOTH skill discovery AND tool execution. A skill delivered over MCP is not an MCP server - it is still a set of instructions the agent reads and follows. The delivery transport and the capability layer are orthogonal.

| Distinction | What it is |
|-------------|-----------|
| SKILL.md content | Procedural instructions for the agent to follow |
| MCP protocol | Transport layer that carries messages |
| MCP server | Executable that provides tools, resources, prompts |
| Skill catalog (MCP-based) | A service that delivers skill instructions over MCP |

## When to Use

Use when:
- The user asks "should this be an MCP server or a skill?"
- You are deciding how to implement a new capability for an agent
- You are evaluating whether an existing MCP server could be replaced by a skill or vice versa
- You need to explain the difference between skills and MCP to a user
- You are comparing agent tool platforms and need to understand the layer boundaries

## Triggers

"should I use MCP or a skill for this", "MCP vs skill", "skill vs MCP server", "difference between skills and MCP", "when to use MCP server", "when to write a skill", "is this an MCP thing or a skill thing", "how do skills relate to MCP", "complementary layers", "skills vs tools", "choose between skill and MCP"

## How Skills and MCP Differ

| Dimension | Skill (SKILL.md) | MCP Server |
|-----------|-----------------|------------|
| What it provides | Procedural knowledge - instructions, methodology, steps | Execution environment - tools, resources, prompts the agent can call |
| When it runs | Read at context-load time, followed as instructions | Runs as a subprocess or remote service, invoked on demand |
| State | Stateless (instructions only) | Can be stateful (databases, file systems, network services) |
| Auth model | None for raw SKILL.md files; delivered through authenticated MCP transport when served from a skill catalog | May need API keys, tokens, or session auth for its tools |
| Discovery | Requires an agent platform (Hermes, OpenClaw) or a catalog (AgentArmory) to index and serve | Discovered via MCP protocol (list tools, read resources) or listed in registries (Smithery, Glama) |
| Modification | Edit the markdown file | Edit the code, rebuild, redeploy |
| Latency | Zero (instructions are in context) | Network/process overhead per call |