Skills, MCP servers, plugins and subagents get talked about as if they were four names for the same feature. They are not. Each one does a different job, each one lives in a different place, and choosing the wrong one is the usual reason a Claude Code setup does not do what its owner hoped. The explainers keep coming, mostly written for developers, with git hooks and code review as the examples. This one uses a sprint report, a retro and a Jira board instead. Every definition below comes from Anthropic's Claude Code documentation, checked on 2 October 2026.
The sentence in Anthropic's own documentation that untangles the most confusion is a short one:
"Skills, subagents, hooks, and MCP servers all work on their own, without a plugin." Anthropic, Claude Code plugins overview
A plugin is the box. The other three are what can go in it. Once that is clear, the rest sorts itself into one question.
One question, four answers
Here is what each one actually is, in the documentation's terms.
A skill is a file of written instructions, named SKILL.md, kept in its own folder. Personal skills live in ~/.claude/skills/ and are available in every project on your machine. Claude can load a skill on its own when the request matches the skill's description, or you can run it by name with a slash, such as /sprint-report. Custom slash commands have been folded into skills, so if you have heard of commands, they are now the same thing.
An MCP server connects Claude Code to another system through the Model Context Protocol, an open standard. Anthropic's examples include Jira, Notion, Asana, Slack and Gmail. A skill tells Claude how to do something. An MCP server gives it something to reach.
A subagent is a separate assistant with its own context window. It starts fresh, does its job, and returns a result to your conversation. Anthropic's documentation is specific about the trade: it does not see your conversation history, the skills you have already run, or the files Claude has already read. That is what keeps a heavy reading job out of your main conversation, and it is also its limit.
A plugin is a package of any of the above, installed as one unit, usually from a marketplace. You need one when you want a whole setup at once, not one piece of it.
"I just want it to write my sprint report. Why do I need four of anything?"
Every guide I open turns into an architecture lesson. I have a report due on Friday and I want a draft of it, not a platform.
Then you need one of the four, and it is a skill. Write down how your sprint report is actually built: the sections, the order, what goes in the summary line, what your sponsor always asks about. Put that in a SKILL.md, give it a one-line description, and run it against last sprint's notes. That is the whole first step.
The other three solve problems you may not have yet. Add them when a specific limit of the skill gets in your way, not before.
"I'm not connecting Claude to Jira. Security will never approve it."
The second anything connects to our ticketing system, someone from security wants a meeting. I would rather not start that fight.
You may not need to, at least at first. Everything in the previous section works with no connection at all: you export or paste the sprint, and the skill does the rest. The connection is a convenience, not a requirement.
When you do want one, the setup is more contained than it sounds. An MCP server can be added at three scopes: just you in one project, everyone in a project, or just you across all your projects. Anthropic's own guidance is blunt about the risk, telling users to verify they trust each server before connecting it, because servers that fetch outside content can expose you to prompt injection. That guidance is worth taking to your security team yourself, before they find it.
"Plugins sound like the easy option. Why not install a few?"
Someone has already built what I need. Why would I write my own skill when I can install a whole plugin in one command?
Sometimes you should. A good plugin from someone who has solved your exact problem is a real shortcut, and Anthropic runs an official marketplace for exactly that. But an installed plugin is not free while it sits there. Anthropic's documentation says an enabled plugin is part of every session, not only the ones where you use it. The names and descriptions of its skills and agents sit in Claude's context on every turn, and those tokens count toward your usage.
It also runs as you. Whatever a plugin does, it does with your permissions. That is a reason to read what a plugin contains before installing it, and to disable the ones you stop using.
The part worth keeping
Four names, four jobs. A skill is how. An MCP server is reach. A subagent is a separate worker with a clean desk. A plugin is the box they ship in. Most setups that disappoint their owners picked by name rather than by job, usually starting with a plugin full of things they did not need.
Start with the job you repeat every sprint, and add the next piece only when the current one runs out.
Sources: Anthropic, skills | Anthropic, MCP | Anthropic, plugins overview | Anthropic, subagents | dev.to, Skills vs MCP connectors vs plugins. Checked 2 October 2026.
