Skills, Connectors, MCP and Plugins: what each one is, and what people actually use them for

Short version

A Skill is an instruction you write down once so you stop retyping it. A Connector is a login that lets the assistant work inside an app you already use — your mail, your drive, your calendar. MCP is the plug shape those connections are built on, and the way to reach an app that has no ready-made connector. A Plugin is a box holding any of the above, so a colleague installs one thing instead of following a setup guide. A Subagent is a second pair of hands for a bulky side-job. Skills teach. Connectors and MCP reach. Plugins hand it over. Subagents take the load.

Three things worth knowing first

A Skill is smaller than you think Anthropic's own published example for writing internal updates is 32 lines of plain English, and contains no code at all See the file: anthropics/skills
A connector can do more than read Connected to Gmail, it can "Send, reply to, and forward emails" — and by default it asks your approval before each one Source: Claude Help Center
The warning is in the vendor's own docs "Anthropic doesn't control what MCP servers, files, or other software are included in plugins and can't verify that they work as intended" Source: Claude Code documentation

Who this is for

Someone who has heard these four words in a meeting, nodded, and gone back to their desk unsure which one fixes the thing that is actually annoying them. No technical background needed. Every term is explained the first time it appears, and every example in this article is something published that you can go and look at yourself.

One month-end, five different problems

Think about the report you put together at the end of every month. It always goes the same way, and it always contains the same five annoyances.

You explain the house format again, because whatever you asked last time came back in the wrong shape. You copy numbers out of the spreadsheet and paste them into the chat window. You need one figure that lives in a system nobody has ever connected to anything. You finish, it works, and then a colleague asks how you did it and the honest answer is a twenty-minute walkthrough. And somewhere in the middle, one bulky step — read all fourteen project updates and find the three that slipped — swallows an hour.

Five annoyances, five different fixes. Nothing here is one product with four names. Anthropic's own documentation sets them out as separate things with separate triggers: you keep typing the same prompt, so save it as a skill; you keep copying data the assistant cannot see, so connect that system; a side task floods your conversation, so hand it to a subagent; a second project needs the same setup, so package it as a plugin (Extend Claude Code).

The annoyance	What fixes it	In one line
"I explain the same thing every single time"	Skill	Write it down once. It follows it from then on.
"It cannot see my mail, my calendar or my files"	Connector	Log in once. It works in them directly.
"The system I actually need isn't on the list"	MCP	The plug shape that lets anything be connected.
"It works for me. Now a colleague needs it too"	Plugin	Put it in a box. They install the box.
"One bulky step swallows an hour"	Subagent	Hand that bit to a second pair of hands.

Read the middle column downward: teach, reach, reach, hand over, delegate. That is the whole distinction. The rest of this article is each row, with examples you can go and look at.

Skill — teaching it how you do something

What it is. A Skill is an instruction you write down once, in ordinary language, so you never type it again. Anthropic's help pages describe skills as folders of instructions that teach the assistant "how to complete specific tasks in a repeatable way", and the examples they give are the unglamorous ones: applying brand style guidelines to documents, following the company's email templates, structuring meeting notes in the company format (Claude Help Center). If you have explained the same thing four times, that explanation wants to be a Skill.

How it works. You save a text file. At the top sits a short description saying when to reach for it; underneath, your instructions. The assistant sees the descriptions all the time so it knows what exists, and reads the full text only when it needs it — which is why a long, thorough one costs you nothing until it is used. The documentation's own advice is to keep that file under 500 lines and push longer reference material into separate files beside it (Agent Skills documentation). No coding is required: "Anyone can create skills by writing instructions in Markdown", and you can also just describe your process out loud and have the assistant write the file for you (Claude Academy).

What one actually looks like. Not one written for this article — this is internal-comms, published by Anthropic in its own public collection, reproduced in full. It is 32 lines, and there is no code in it anywhere (see the file):

A real, published Skill for writing internal communications, quoted in full from Anthropic's public collection. Note what it is: a description saying when to use it, a list of what it covers, and three instructions.

---
name: internal-comms
description: A set of resources to help me write all kinds of
  internal communications, using the formats that my company
  likes to use. Claude should use this skill whenever asked to
  write some sort of internal communications (status reports,
  leadership updates, 3P updates, company newsletters, FAQs,
  incident reports, project updates, etc.).
---

## When to use this skill
To write internal communications, use this skill for:
- 3P updates (Progress, Plans, Problems)
- Company newsletters
- FAQ responses
- Status reports
- Leadership updates
- Project updates
- Incident reports

## How to use this skill

To write any internal communication:

1. **Identify the communication type** from the request
2. **Load the appropriate guideline file** from the `examples/`
   directory
3. **Follow the specific instructions** in that file for
   formatting, tone, and content gathering

If the communication type doesn't match any existing guideline,
ask for clarification or more context about the desired format.

Two things are worth noticing. The description does not describe the Skill — it describes when to reach for it, and it lists the situations by name: status reports, newsletters, FAQs, incident reports. That naming is what makes the assistant recognise the moment. And the Skill holds almost no content of its own: it points at four separate files, one per type of update, so the long material is only read when that particular type comes up.

Published Skills you would recognise

Every one of these is in the same public collection, and each is one folder you can open (anthropics/skills):

pdf — the file-wrangling one. Its own description covers "combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks", filling in PDF forms, and running text recognition over a scanned document so it becomes searchable.
docx, xlsx, pptx — Word documents, spreadsheets and slide decks. The spreadsheet one is written for "cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets", which is a Tuesday afternoon most people would recognise.
brand-guidelines — the published one applies Anthropic's own official colours and typography to whatever is being produced. It is there as the shape to copy for your own house style, so nobody has to remember the colour codes.
doc-coauthoring — walks you through writing a proposal or a spec in three stages: gathering context, refining the structure, then testing it on a reader.
slack-gif-creator — makes an animated GIF sized correctly for Slack. Trivial, and a perfect illustration of the scale these things work at.

Who it is for. Anyone who repeats themselves. It needs no technical background — it is a text file with instructions in it, and you can dictate the first one to the assistant rather than writing it.

Where it stops. Three limits, and the last one is the one that surprises people.

A Skill only knows what you put in it. It cannot open your mail, look in your files or check a system. It is knowledge, not access.

If the description at the top is vague, nothing happens at all. The documentation is blunt about this: "If descriptions are vague or overlap, Claude may load the wrong skill or miss one that would help" — so a Skill that appears broken is usually a Skill whose top line does not say when to use it (Extend Claude Code).

And a Skill is a request, not a rule. The same documentation says it plainly: an instruction like "never edit this file" written into a Skill "is a request, not a guarantee". If something must hold every single time, a Skill is the wrong tool for it (Extend Claude Code).

Connector — letting it into your apps

What it is. A Connector is a login. Anthropic's own definition: connectors "let Claude access your apps and services, retrieve your data, and take actions within connected services" (Claude Help Center). You connect it once, and from then on it works inside that app instead of you copying things across.

How it works. In the sidebar, "Click 'Customize' in the left sidebar", pick Connectors, press the plus button and log in — the same permission screen you have seen a hundred times (Claude Help Center). The apps on the list are the ones most offices already run: Gmail, Google Calendar, Google Drive, Slack, Microsoft 365, Notion, Linear, Asana, Jira, Confluence, GitHub, Figma, Stripe, PayPal and HubSpot all appear by name in Anthropic's own pages (Claude Academy).

Questions people actually ask

These are not invented. They are the example questions Anthropic publishes in its own training course for connectors (Claude Academy):

"What did the team decide about the timeline in yesterday's discussion?" — the chat-app question. Reading forty messages to find the one decision is nobody's idea of work.
"Find the email thread where we discussed the vendor contract." — the one that replaces ten minutes of search-box guessing.
"What are my highest priority tasks due this week?" — asked of the project tracker, without opening it.
"Summarize the meeting notes from last week's product review." — asked of the notes app.
"What does our style guide say about using contractions?" — the documentation question, which is the one people never think of and use most once they have it.

The value is rarely one app. It is that the month-end question — what happened, what was decided, what is still open — is a calendar question, a mail question and a files question at the same time, and a connector lets you ask it once.

Who can switch it on. On a personal plan, you. In a company on a Team or Enterprise plan, "an Owner or Primary Owner needs to enable them for the organization" first — so if the option is missing, it is not broken, somebody just has not turned it on (Claude Help Center).

Where it stops. Four limits, and the first two are usually stated wrongly.

It can act, not only read — and it asks first. Connected to Gmail it can "Send, reply to, and forward emails", and connected to Google Calendar it can "Create, update, and delete events". The safety is not that it refuses; it is that "By default, Claude asks for your approval before each of these actions". You can also narrow that yourself: the permissions screen lets you "Allow Claude to search and summarize email, but prevent it from sending messages" (Google Workspace connectors).

It can only see what you can see. "Claude mirrors your existing permissions—you cannot access information you don't already have access to in Google Workspace." It runs on your account, so it cannot open a file or a channel you could not open yourself (Google Workspace connectors).

Attachments are the gap nobody expects. With Gmail, "Attachment content is not directly accessible" — it sees that the attachment is there, not what is inside it. With Drive, it "extracts text content only" and "Images embedded in documents are not processed". If your month-end numbers arrive as a picture in a PDF, that is the sentence that ruins the plan (Google Workspace connectors).

And on the free plan you get one. "Free users are limited to one custom connector", so trying three at once is not a free experiment (Claude Help Center).

MCP — the plug shape

What it is. MCP — the Model Context Protocol — is an agreed shape for the plug between an assistant and an app. Its own documentation puts it better than any paraphrase: "Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems" (modelcontextprotocol.io). It is not a feature. It is a shape everybody agreed on, and it is open, so it is not one company's property: Claude, ChatGPT, VS Code and Cursor all speak it.

You are already using it. The connectors in the previous section are built on this shape — the ready-made ones are simply the plugs somebody already made for you. The word only starts to matter to you when the system you need is not on the list: an industry system, something old, something built in-house. Then MCP is the answer to "can we connect ours?", and the answer is usually yes, as a piece of work for whoever looks after that system.

What gets connected this way

The apps that already have one waiting. Notion, Asana and Airtable each connect with a single command in Anthropic's own documentation — one line, no integration project (MCP documentation).
Your own database, read-only. The most useful one for most offices: asking questions of a system nobody remembers how to query. The documentation's own instruction is to "Use a read-only database user in the connection string so the queries Claude runs can't modify data" — so a question can never accidentally become a change (MCP documentation).
A design file, turned into a working page. The protocol's own list of what it makes possible includes "Claude Code can generate an entire web app using a Figma design" (modelcontextprotocol.io).
Your calendar and your notes at the same time. Also from the same list: "Agents can access your Google Calendar and Notion, acting as a more personalized AI assistant" (modelcontextprotocol.io).
And, genuinely, a 3D printer. "AI models can create 3D designs on Blender and print them out using a 3D printer." It is on the official page. It is worth knowing how wide the shape is (modelcontextprotocol.io).

Where it stops. Two things, before somebody enthusiastically connects nine of them.

More is not better. The vendor's own words: "Every feature you add consumes some of Claude's context. Too much can fill up your context window, but it can also add noise that makes Claude less effective; skills may not trigger correctly, or Claude may lose track of your conventions." In plain terms: a machine wired to everything is vaguer than one wired to the three or four things you actually use. Keep a handful (Extend Claude Code).

The shape is not a safety promise. MCP describes how the plug fits, not whether the thing on the other end is trustworthy. The documentation is direct: "Verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection risk." Prompt injection means text arriving from outside — inside a web page, a document, a ticket — that is written to look like an instruction to the assistant, so the assistant follows it as though you had typed it. Which leads to the section below (MCP documentation).

Plugin — the box you can hand someone

What it is. A Plugin is a box. It can hold Skills, connections and small automations together, and one person sets it up so everybody else installs a single thing. The documentation calls it exactly that: "the packaging layer. A plugin bundles skills, hooks, subagents, and MCP servers into a single installable unit" (Extend Claude Code).

How it works. The difference from a Skill is not what it does — it is that it can be handed over. Before you install one you get a details screen showing three useful things: how much room it will take up every turn, when it was last updated, and a list of exactly what it will add. Then you choose whether it is installed for you across everything, for one project, or for everyone working on that project (Discover and install plugins).

Plugins that exist today

These are all in Anthropic's own official catalogue, named on its documentation page (Discover and install plugins):

The apps you already pay for. One plugin each for github, gitlab, slack, notion, linear, asana, atlassian (that is Jira and Confluence), figma, sentry (error alerts), vercel, firebase and supabase. Each one bundles the connection so nobody has to set it up by hand.
commit-commands — the small, boring one people install first: it packages the routine of saving your work and writing up what changed.
security-guidance — reviews each change as it is made and tells the assistant to fix what it finds, in the same sitting rather than in a review three days later.
pr-review-toolkit — several reviewers at once, each looking for something different, instead of one pass that catches one kind of problem.
explanatory-output-style and learning-output-style — these change how the assistant talks to you rather than what it can do: one explains why it made each choice, the other turns the session into something closer to a lesson. A good reminder that a plugin does not have to add a capability to be worth having.

And the one nobody publishes because it is yours: the box holding your own report format, your own invoice check and your own house style. Written once, installed by everyone.

Where it stops. Two limits, and the first is the reason the next section exists.

A plugin is exactly as trustworthy as whoever made it. Anthropic states it without softening: "Plugins and marketplaces are highly trusted components that can execute arbitrary code on your machine with your user privileges. Only install plugins and add marketplaces from sources you trust" (Discover and install plugins).

And installing more of them is not free. The tool itself now nudges you about this: it lists plugins you have not used "in at least two weeks, over a span of at least 10 sessions" under a heading called Not used recently, precisely so you can find "plugins that still add startup and context cost even though you no longer use them". If the software is telling you to prune, prune (Discover and install plugins).

Subagent — a second pair of hands

What it is. A separate worker handed one self-contained job, which does it out of sight and comes back with the answer rather than the working. The documentation describes exactly when to reach for one: "Use one when a side task would flood your main conversation with search results, logs, or file contents you won't reference again: the subagent does that work in its own context and returns only the summary" (Subagents documentation).

How it works. It gets its own clean workspace and its own instructions, and only the summary comes back to you. That is the whole point — the fourteen project updates never land in your conversation, only the three that slipped. The named examples in the documentation are the obvious shapes: a code-reviewer, a test-runner, a browser-tester, a read-only database reader, and two built-in ones called Explore and Plan for "read a lot, report back briefly" and "think it through before touching anything" (Subagents documentation).

Who it is for. Anyone whose job contains a step where the work is bulky, the answer is short, and you do not need to watch it happen. Reading fifty documents to find three. Running the same check across nine folders.

Where it stops. Because it starts in a clean room, it knows nothing you have not told it: one "that isn't a fork starts fresh and may need time to gather context", so a vague instruction produces vague work faster than before. It also lives and dies with the conversation — "Subagents work within a single session" — so it is not a background job you can walk away from for a week. The upside worth knowing: you can point the bulky, dull ones at a cheaper model, which is the documentation's own suggestion for keeping the bill down (Subagents documentation).

Which one do you actually need?

Find the sentence you have said out loud and read across. Most of these triggers come straight from Anthropic's own "when to add what" table, which is worth reading in full if you want the developer-facing version (Extend Claude Code).

If you catch yourself saying…	You want	Where to start
"I explain the same thing every time"	A Skill	Describe it out loud and let it write the file.
"It never does it the way we do it here"	A Skill	Same file. Name the situations it applies to.
"I keep copying things out of my mail and files"	A Connector	Customize, Connectors, plus, log in.
"What happened while I was away?"	A Connector	Connect calendar, drive and mail — the answer needs all three.
"Our system isn't on the list"	MCP	A conversation with whoever runs that system.
"The numbers are in a scanned PDF attachment"	Not a connector	It reads the attachment's name, not its contents. Get the file out first.
"This must happen every time, no exceptions"	Not a Skill	A Skill is a request, not a guarantee. You need a rule enforced outside it.
"A colleague needs what I've got"	A Plugin	Put it in a box; they install the box.
"One bulky step blocks the rest"	A Subagent	Hand off that step alone, and ask for the summary.
"It's got slow and vague since we set it all up"	Fewer of everything	Turn off what you do not use. Keep a handful.

The same month-end, with each step labelled by what handles it.

Before you install anything

This is the short section that matters most, and it is not technical.

A plugin or a connection is software written by someone else, running with your permissions. The vendor says so in two places without any softening. On plugins: they "can execute arbitrary code on your machine with your user privileges. Only install plugins and add marketplaces from sources you trust." And on what has and has not been checked: "Anthropic doesn't control what MCP servers, files, or other software are included in plugins and can't verify that they work as intended" (Discover and install plugins).

That is not a reason to avoid them. It is a reason to know which shelf you are taking something off. There are two official catalogues and they are not the same: one is "curated by Anthropic, and inclusion is at Anthropic's discretion", while the community one holds "third-party plugins that have passed Anthropic's automated validation and safety screening", with each one "pinned to a specific commit SHA" — meaning you get the exact version that was screened, not whatever the author pushed last night. Automated screening plus a frozen version is a real filter. It is not a person reading the code (Discover and install plugins).

Three habits that cost nothing. Read the "will install" list on the details screen before you press install — it is there so you can see exactly what you are adding. Connect any database with a read-only account, so "the queries Claude runs can't modify data" (MCP documentation). And leave the approval prompts on for anything that sends, deletes or shares — the default is that it asks first, and turning that off to save three seconds is a bad trade (Google Workspace connectors).

Bottom line

Three verbs separate the four words. A Skill teaches — the thing you keep explaining, written down once, read only when it is needed, so writing it out properly costs you nothing (Agent Skills documentation). A Connector and MCP both reach — the first for apps somebody already wired up, the second for everything else. A Plugin hands it over, so what works for you works for a colleague in one step. A Subagent takes the bulky job off your desk.

Nearly everyone should start in the same place, and it is the cheapest one: take the thing you explain most often and turn it into a Skill. You do not have to write it — describe your process out loud and have the assistant produce the file (Claude Academy). Anthropic's own published example is 32 lines of ordinary English. That is the whole barrier to entry.

Three things to keep in your head afterwards. A Skill is a request, not a rule, so anything that must hold every time needs enforcing somewhere else. A connector can act, not just read — it can send that mail, and the thing standing between you and a mistake is the approval prompt you were about to switch off. And more is worse: the vendor's own documentation says extra features "can also add noise that makes Claude less effective", and the software now lists the ones you have stopped using so you can remove them (Extend Claude Code).

FAQ

What is the difference, in one line each?

A Skill is an instruction you wrote down. A Connector is a login to an app you already use. MCP is the plug shape underneath connectors, and the way to connect something that has no ready-made one. A Plugin is a box holding any of those so someone else installs it in one step. A Subagent is a second pair of hands for a bulky side-job.

How is a Skill different from just telling it what I want?

Nothing, the first time. The difference is the fourth time, and the difference when a colleague needs the same result. Anthropic's help pages give the everyday cases: applying brand guidelines to documents, following company email templates, structuring meeting notes in the house format (Claude Help Center).

Do I need to be technical to write one?

No. "Anyone can create skills by writing instructions in Markdown—no coding required for simple skills." You can also skip the writing entirely and describe your process in conversation, which "makes Skills accessible to anyone, regardless of technical background" (Claude Academy).

Will a long Skill slow things down?

No. Only the short description is in view all the time; the body loads when it is used. The documentation's guidance is to keep the main file under 500 lines and move longer reference material into separate files next to it (Agent Skills documentation).

I wrote a Skill and it seems to be ignored. Why?

Almost always the description at the top. "If descriptions are vague or overlap, Claude may load the wrong skill or miss one that would help." Name the situations it applies to, in the words you would actually use (Extend Claude Code).

Can it see things I am not allowed to see?

No. "Claude mirrors your existing permissions—you cannot access information you don't already have access to in Google Workspace." It works through your own account (Google Workspace connectors).

Can it send email for me?

Yes — this is the one most explainers get wrong. Connected to Gmail it can "Send, reply to, and forward emails", and "By default, Claude asks for your approval before each of these actions". If you would rather it never sent anything, the permissions screen lets you allow searching and summarising while blocking sending (Google Workspace connectors).

Why can't it read the attachment I sent it?

Because with Gmail, "Attachment content is not directly accessible" — it sees the attachment exists, not what is in it. Drive files give up their text but not their pictures: "Images embedded in documents are not processed" (Google Workspace connectors).

The Connectors option isn't there. Is it broken?

Probably not. On Team and Enterprise plans "an Owner or Primary Owner needs to enable them for the organization" before anyone can connect anything (Claude Help Center).

Do I have to pay to try a connector?

Not to try one. "Free users are limited to one custom connector", so you can test the idea before deciding whether the whole setup is worth it (Claude Help Center).

Why does everything feel slower since we connected lots of apps?

Because each thing you add takes up room before you have asked anything: "Too much can fill up your context window, but it can also add noise that makes Claude less effective." The tool now flags plugins you have not touched in two weeks so you can remove them (Discover and install plugins).

Is it safe to install a plugin somebody sent me?

Treat it as software from a stranger, because that is what it is. Plugins "can execute arbitrary code on your machine with your user privileges", and Anthropic "doesn't control what MCP servers, files, or other software are included in plugins". Prefer the official catalogue, and read the "will install" list before pressing install (Discover and install plugins).

What is prompt injection, and should I care?

It is text arriving from outside — in a web page, a document, a support ticket — written to look like an instruction to the assistant, so it gets followed as though you had typed it. It matters most for connections that fetch outside content: "Servers that fetch external content can expose you to prompt injection risk" (MCP documentation).

Is MCP a Claude thing?

No, it is open, which is the point of it. "AI assistants like Claude and ChatGPT, development tools like Visual Studio Code, Cursor, MCPJam, and many others all support MCP" — so a connection built once is not tied to one assistant (modelcontextprotocol.io).

Do we need all four?

Most people need one. A Skill fixes repeating yourself, which is the usual complaint. Connectors follow once you notice how much time goes into copying things across. MCP and Plugins matter when a real system has to be reached, or when what works for you has to work for everyone.

Where should I start on Monday?

With the thing you explain most often. Describe it out loud, let the assistant write it into a file, and use it for two weeks. Then look at what is still annoying you — that tells you which of the other four you actually need.

ISA After Hours · Augsburg

Setting one of these up and want a second opinion?

ISA After Hours is a community of Israeli and international tech professionals in Augsburg. We run these tools on our own work and compare what actually held up — which finds the limits faster than reading another guide.

Join ISA After Hours →

Sources

Every page below was opened and read while writing this article, and every quotation is copied from it. Primary documentation first.