# Tools — three shapes, one purpose

> **What you will learn**
> What a tool is, why it arrives under three names — **skill, plugin, MCP** —
> why that distinction is not the user's problem, and the **three real
> differences** you do have to know anyway.

## A tool is "something it can do"

You saw the structure in [Tool use](/guide/ai-tool-use). A model only produces
text, but hand it **a list of actions it can call** and it picks one and calls it.

```
Without tools:  "You may wish to check the stock level"   ← it only talks
With tools:     checkStock(sku="A-102") → 37              ← it actually looks
```

That is all a tool is. **An action with a name and a description attached, so the
model can call it.** Nothing more.

## But there are three names

Go to add one and you meet three words. All three amount to "give the work more
things it can do"; the packaging differs.

| | What it holds | Where it lives | Scope |
|---|---|---|---|
| **Skill** | A procedure written down | `.claude/skills/` | This working folder |
| **Plugin** | A bundle of several | `~/.claude/plugins/` | **The whole machine** |
| **MCP server** | A protocol for talking to an outside program | `.mcp.json` | This working folder |

```mermaid
graph TD
  A["I want the work to be<br/>able to do more"] --> B["Skill<br/>write down the procedure"]
  A --> C["Plugin<br/>install a bundle"]
  A --> D["MCP server<br/>attach an outside program"]
  B --> E["Result: more callable actions"]
  C --> E
  D --> E
```

## The distinction belongs to whoever built it

This is the point of the chapter. **Which of the three something is was decided
by its author; it is not a question the user should have to answer.**

```
kordoc  → a CLI installed via npm, and an MCP server, and a plugin
docx    → a skill, which calls npm from inside itself
```

One thing carries three names at once. So **"which kind was the one I'm looking
for?" is usually a question with no answer.**

> **Kind should be a badge, not a page.**

The HyperTeams "Tools" screen works that way. It used to be three tabs — skills,
plugins, MCP — and they were folded into one, because the user is trying to do
one thing (**"what should I let this work do?"**), and standing tabs in front of
that only **adds a question they have to answer first.**

Plenty of mature tools reached the same answer in the same place.

```
VS Code extensions · Raycast · Obsidian community plugins
  → all of them: one list of what's installed + one place to add
  → kind is a small badge on the row
```

**Which is why this part does not open with MCP.** MCP is one of the three
shapes, and it gets its own chapters further down.

## But the handling is not the same

You can ignore the packaging right up until you touch it. Then the three differ.
These are worth memorising.

| | Can be toggled off | Deleting removes it from |
|---|---|---|
| Skill | **No** | This working folder |
| Plugin | No (per-folder use is separate) | **The whole machine** |
| **MCP server** | **Yes** — the only one | This working folder |

Each row burns you once.

- **"Cannot be turned off" and "is turned off" are different.** A skill has no
  temporary off switch. Not wanting it means deleting it.
- **Deleting a plugin removes it from the machine, not the folder.** You just
  took away something other work was using.
- **Not everything you can see is yours to delete.** Only what was put into this
  folder; rows that came from the machine-wide install or from policy cannot be
  touched here.

## Questions to ask of any one tool

| Question | Why | Where |
|---|---|---|
| **What does it do?** | The model picks using only the description | [What an MCP server gives you](/guide/mcp-capabilities) |
| **What permission is it asking for?** | Attaching one is granting permission | [Tools are permissions](/guide/mcp-security) |
| **Where does it install?** | This is where the cost of undoing splits | The table above |

The third is the easy one to skip. **Folder-level is cheap to undo; machine-level
is expensive.** If you are only trying something out, start on the folder side.

## Where you actually touch this

| Where you work | The place |
|---|---|
| HyperTeams on your machine | The "Tools" tab — all three shapes in one list |
| A Connect workspace | [Connecting an MCP server](/guide/cn-mcp) · [Turning tools on and off](/guide/cn-tool-settings) |
| Capturing a procedure in writing | [Capturing procedures as skills](/guide/ht-skills) |

---

## Check

**1. Why not split skills, plugins, and MCP into tabs by kind?**

<details>
<summary>Answer</summary>

**Because which one something is was the author's decision.** kordoc is a CLI and
an MCP server and a plugin; docx is a skill that calls npm inside itself. Tabs
force the searcher to answer "which kind was mine?" **first** — usually a question
with no answer. A badge is enough.
</details>

**2. Which of the three can be toggled off? What do you do with the others?**

<details>
<summary>Answer</summary>

**Only MCP servers.** Skills and plugins have no temporary off switch, so not
using one means deleting it. **"Cannot be turned off" and "is turned off" are
different.**
</details>

**3. What is the particular hazard when deleting a plugin?**

<details>
<summary>Answer</summary>

**It disappears from the whole machine, not this working folder.** You have taken
away something other work was using. Skills and MCP servers are folder-level, so
their blast radius is narrower.
</details>

---

Whatever the shape, adding them leads to the same problem →
[When you have too many tools](/guide/mcp-too-many-tools)