Back to articles
MCP & plugins explained•October 29th, 2026

Claude Code plugins vs skills vs MCP servers, explained

M

MechaBee Content Team

7 min read
Claude Code plugins vs skills vs MCP servers, explained

You install a plugin and get a new slash command. You add an MCP server and get new tools. You write a skill and Claude follows a new procedure. All three extend what your agent can do, and the names overlap enough that it's easy to pick the wrong one.

This post sorts out Claude Code plugins vs skills vs MCP servers: what each one is, what it adds, how they fit together, and how the same three layers work in Codex. Then we take a real plugin apart.

The short answer

SkillMCP serverPlugin
What it isInstructions, knowledge or a workflow, written in MarkdownA server that gives the agent tools and data from an outside systemA package that installs skills, MCP servers and other parts as one unit
What it addsWhat the agent knows and how it worksWhat the agent can reach and doA way to share a setup
Where it livesA SKILL.md fileA server, configured in an MCP config such as .mcp.jsonA folder with a manifest, such as .claude-plugin/plugin.json
ExampleYour deploy checklist, run as /deployQuery a database, post to Slack, read a project trackerA team's review skill plus the MCP server it relies on

In one sentence: skills are knowledge, MCP servers are tools, plugins are packaging. Most real setups use all three.

Skills: what your agent knows and how it works

In Claude Code, a skill is a Markdown file containing knowledge, workflows or instructions. You can run it yourself with a command like /deploy, or Claude loads it on its own when your request matches its description. Only the description sits in context all the time; the full skill loads when it's used. (Extend Claude Code, checked 29 October 2026.)

Skills come in two kinds:

  • Reference skills hold knowledge the agent uses throughout a session, such as your API style guide.
  • Action skills tell the agent to do something specific, such as running your release checklist.

A skill doesn't give the agent any new abilities. It can only use the tools the agent already has. What it changes is judgement: which steps to take, in what order, and what to avoid.

MCP servers: the tools and data your agent can reach

The Model Context Protocol (MCP) is the standard way to connect an agent to an outside system. An MCP server offers a set of tools, such as "search issues" or "read file", and the agent can call them. The server handles the connection and the sign-in. In Anthropic's words, MCP "connects Claude to external services", while skills extend what Claude knows, including how to use those services well.

How is MCP different from an API? An API is a service's own interface, designed for programmers. An MCP server describes its tools in a form any MCP-compatible agent can discover and call, often by wrapping an API underneath. One server works in Claude Code, Codex and other MCP clients.

Do you need an MCP server for Claude Code? Not for most coding. Claude Code's built-in tools read and edit files, search, run commands and browse the web. Add an MCP server when you keep copying information from a system the agent can't see, such as a ticket tracker, a database or your team's documents.

Plugins: how skills and MCP servers are packaged and shared

A Claude Code plugin is a folder of components that Claude Code installs and loads as one unit. Its manifest lives at .claude-plugin/plugin.json, and it can hold skills, subagents, hooks and MCP servers. Plugin skills are namespaced, like /my-plugin:review, so several plugins can sit side by side. Most plugins come from a marketplace, a catalogue that lists plugins and where to fetch them. (Plugins overview, checked 29 October 2026.)

So are Claude Code plugins just skills? No. A plugin can contain only a skill, but its job is packaging: it delivers whatever a setup needs (skills, the MCP servers they depend on, hooks, subagents) in one install, at one version, to everyone who uses it.

The official guidance puts it as a trigger: when a second repository needs the same setup, package it as a plugin.

Skills vs plugins in Codex: the same three layers

Codex uses the same model. In OpenAI's docs, a skill packages instructions and supporting resources for a specific task, and a plugin is an installable bundle that can include skills and MCP servers, plus lifecycle hooks in Codex. In Codex you mention a skill with $. (Skills & Plugins, checked 29 October 2026.)

The packaging differs in detail. A Codex plugin can use a portable plugin.json at its root, and the older .codex-plugin/plugin.json layout is still supported. (Package your plugin, checked 29 October 2026.) The idea is the same: skills in a skills/ folder, MCP servers in an MCP config, and a manifest that ties them together.

Because MCP is a shared standard, the same MCP server works for both. That's what makes one plugin folder that installs in both tools possible.

A real plugin, taken apart

MechaBee's Keeper Workspace Content plugin lets Claude Code or Codex save and look up work in MechaBee: read and revise workspace documents, read what a Keeper app shows, and carry out the tasks an app offers agents. Here's how it's built.

keeper-workspace-content/
├── .claude-plugin/plugin.json   ← manifest for Claude Code
├── .codex-plugin/plugin.json    ← manifest for Codex
├── .mcp.json                    ← one remote MCP server
└── skills/
    └── keeper-workspace-content/
        └── SKILL.md             ← one skill

The MCP server gives the tools. .mcp.json points to one server over HTTP, with sign-in through MechaBee (trimmed):

{
  "mcpServers": {
    "workspace": {
      "type": "http",
      "url": "https://mechabee.com/mcp/workspace",
      "auth": "oauth"
    }
  }
}

The server offers tools such as listing workspaces, reading a file and writing it back. Every call is limited to what the signed-in person can reach.

The skill gives the judgement. SKILL.md tells the agent when to use those tools and how to use them safely. For example: read a document before editing it, and send the version you read with the edit, so a concurrent change produces a conflict instead of a silent overwrite. On a conflict, reread and merge; never retry blindly. The server could enforce the version check on its own, but only the skill teaches the agent what to do when it fails.

The plugin makes it one install. The same folder carries both manifests, so it installs in Claude Code and in Codex, and the skill and the server always arrive together at matching versions.

Our other plugin, Keeper App Author, follows the same shape: one skill that knows how to design and change Keeper apps, and one MCP server that validates and deploys them.

Which one do you need?

Start small and add a layer when a specific need comes up:

  • You keep pasting the same procedure into chat. Write a skill.
  • You keep copying data from a system the agent can't see. Connect an MCP server.
  • A skill only works if a certain MCP server is connected. Ship them together in a plugin.
  • A second repository, or a teammate, needs the same setup. Package it as a plugin and share it through a marketplace.
  • A rule must hold every time, without the agent deciding. That's a hook, not a skill.

If a project rule should always apply, such as "use pnpm, not npm", put it in your CLAUDE.md (or AGENTS.md) rather than a skill.

FAQ

What's the difference between skills, connectors and plugins in Claude? A connector is an MCP server you add from Anthropic's directory; the docs say directory connectors use the same MCP infrastructure as Claude Code. Skills and plugins are as described above.

What's the difference between a plugin and a skill in Codex? The same as in Claude Code: a skill is a packaged procedure; a plugin is an installable bundle of skills and MCP servers.

Should I use Claude Code skills? Yes, once you notice yourself repeating a procedure. They're the lightest way to make an agent follow your process.

Can one plugin work in both Claude Code and Codex? Yes, if it carries a manifest for each and its MCP servers are reachable from both, as in the example above.

Summary

Skills teach your agent how to work, MCP servers connect it to the systems it needs, and plugins package both so a whole team gets the same setup. Claude Code and Codex use the same three layers, so one well-built plugin can serve both.

To see a working example, install the MechaBee plugins in Claude Code or Codex and sign in once. You can start free with included credits.

Back to articles