Skill vs Plugin vs MCP

Three terms get used interchangeably and shouldn't be. Here's the clean mental model, with one common mistake corrected up front.

The correction: a plugin is NOT an MCP. A plugin is a bundle of skills (and optionally agents / commands). An MCP is a capability connection (tools, APIs, auth). They sit on different layers — a plugin can ship skills that use an MCP, but the plugin is not the MCP. Older notes conflated "plugins/MCP = capabilities"; that's the conflation this page fixes.

The three layers

🧠 Skill — the intelligence

A workflow, expertise, or procedure written as SKILL.md + optional supporting files. It's what to do and how. Lazy-loaded: ~200 tokens at startup, full body loaded only when its description matches the task.

🔌 MCP — the capability

A connection to external tools, APIs, data, or execution (Gmail, a database, a browser). It's what's possible. Eagerly loaded — every tool's schema enters context at startup (no progressive disclosure), which is why heavy MCPs are expensive.

📦 Plugin — the bundle

A packaged collection of skills (and possibly agents, commands, hooks, or MCP configuration) distributed as one unit. It's how you ship a set of capabilities together. A plugin is a delivery wrapper, not a runtime thing.

How they relate

Skills orchestrate MCPs — a skill says "run my morning digest" and handles the 3–4 MCP tool calls underneath; the user never sees the tool names. A plugin packages one or more such skills so they install together.

	Skill	MCP	Plugin
Layer	Intelligence	Capability	Distribution
Answers	How to do it	What's possible	How to ship a set
Loading	Lazy (~200 tok)	Eager (all schemas)	n/a (install-time)
Unit	`SKILL.md` + files	Server + tools	Bundle of skills (+agents)
Example	"Morning digest"	Gmail / M365 connector	A client's whole skill pack

Skills ≡ Commands (a settled point)

In current Claude Code, custom slash commands and skills are the same mechanism. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way. The only practical difference: skills get a supporting-files directory; legacy commands don't. Turning model-invocation off makes a skill behave purely as a slash command. So "that's a command, not a skill" is a distinction that doesn't exist at the platform level.

The rule of thumb

Need logic / a workflow	Write a skill
Need to reach an external system	Connect an MCP
Need to hand someone a set of skills at once	Ship a plugin

Source of truth seeded from _brain/knowledge/skills/skills.md (Skills vs Plugins; Skills and Commands merged, 260504; Skill vs MCP loading asymmetry, 260504), corrected for the stale plugin≡MCP conflation per planning/DECISION-TREE.md §D7.2. Part of the Skill Viewer rung-1 build.