Direct Answer: What is the Difference Between Plugins, Skills, and MCP Servers?
In Claude Code, Skills, MCP Servers, and Plugins operate at three different layers of the assistant architecture:
- Skills (
SKILL.md) provide procedural knowledge (Layer 1 primitive). They are markdown prompt packages that teach the AI how to think and act contextually. - MCP Servers (
.mcp.json) provide tool and data connectivity (Layer 1 primitive). They are protocol bridges connecting Claude to external databases, APIs, or local system tools. - Plugins (
.claude-plugin/plugin.json) provide the distribution and packaging layer (Layer 2). A plugin bundles skills, slash commands, dedicated subagents, event hooks, and MCP servers into a single versioned unit that can be installed with one command.
Architectural Comparison Table
| Capability | Agent Skill | MCP Server | Claude Code Plugin |
| --- | --- | --- | --- |
| Primary Purpose | Contextual workflows & rules | Tool execution & data integration | Multi-component distribution |
| Manifest / File | SKILL.md | .mcp.json | .claude-plugin/plugin.json |
| Execution Trigger | Loaded contextually by LLM | Invoked by tool call | Loaded at session startup / /reload-plugins |
| Can Bundle Other Tools? | No | No | Yes (Skills, Commands, Agents, Hooks, MCPs) |
| Enforces Deterministic Guardrails? | No (probabilistic LLM prompt) | No (tool interface) | Yes (via hooks/) |
| Installation Method | File copy or git clone | Add to .mcp.json | /plugin install <name>@<market> |
| Multi-machine Distribution | Manual file copying | Manual config entry | Automated via /plugin marketplace |
Deep Dive: Layer 1 Primitives
1. Agent Skills (Layer 1: Knowledge)
An agent skill is fundamentally an educational resource for the LLM. Located in.claude/skills/<name>/SKILL.md, it contains instructions on formatting code, reviewing pull requests according to corporate guidelines, or orchestrating complex multi-file refactors.
- Strengths: Lightweight, human-readable, requires zero runtime dependencies.
- Limitations: Dependent on the model's instruction-following adherence; cannot deterministically prevent an unwanted action.
2. Model Context Protocol Servers (Layer 1: Tooling)
MCP is an open standard created by Anthropic that allows assistants to communicate with external applications over JSON-RPC. Configured in.mcp.json, an MCP server exposes callable tools (e.g., github_create_issue, query_postgres_db).
- Strengths: Grants the AI live access to company databases, cloud systems, and third-party APIs.
- Limitations: Exposes functional capabilities but does not dictate how or when the agent should invoke them.
Deep Dive: Layer 2 Packaging (Plugins)
A Claude Code plugin does not replace Skills or MCP servers—it packages them together into a unified workflow.
Consider a complete Git automation workflow. Without plugins, an engineering team would have to:
- Copy a
git-commitskill into~/.claude/skills/. - Write a
/prcommand in~/.claude/commands/. - Add a GitHub MCP server to
.mcp.json. - Configure safety hooks in
settings.jsonto block force pushes.
With the plugin architecture, an author bundles all four items into a single repository:
git-suite-plugin/
├── .claude-plugin/plugin.json
├── skills/git-commit/SKILL.md
├── commands/pr.md
├── hooks/block-force-push.sh
└── .mcp.jsonA developer on the team simply runs:
/plugin install git-suite@company-marketplace --scope projectAll four components are immediately registered, version-controlled, and synchronized across every developer working on the repository.
Which One Should You Build?
- Build a Skill if you are documenting internal team conventions, design system guidelines, or repeatable prompt instructions that do not require custom shell hooks or new slash commands.
- Build an MCP Server if you need to connect Claude Code to an internal corporate API, database, or hardware device.
- Build a Plugin if you have created a complete workflow combining skills, slash commands, security hooks, or MCP tools that you want other developers or the open-source community to install with one command.
Frequently Asked Questions
Can a plugin include an MCP server?
Yes. Plugins can include an.mcp.json file in their root directory. When the plugin is installed and enabled, Claude Code automatically registers the declared MCP servers within the plugin's scope.
Can I convert an existing skill into a plugin?
Yes. Create a.claude-plugin/plugin.json manifest file at your repository root, place your skill folder inside a skills/ directory, and define a name and version. The repository is then immediately installable as a plugin.