If you have been following the AI space, you have probably seen the words tools, skills, plugins, MCP, hooks, and actions used interchangeably — sometimes by the same company in the same article.

They are not the same thing. But they are all related, and the confusion is understandable because the industry has not agreed on standard names yet.

Here is a plain-language guide to what each one actually means. (If you want a deeper look at how skills work in practice, see our companion post: How AI Agents Use Skills to Get Things Done.)

Start With the Simplest Idea: a Tool

A tool is a specific task that an AI is allowed to request.

That is it. Behind the scenes it looks like a short description card:

{
  "name": "search_cases",
  "description": "Search legal cases by keyword and location",
  "needs": {
    "keyword": "what to search for",
    "location": "SG, MY, or HK"
  }
}

When a company builds an AI assistant, they define a list of tools up front. The AI can only request things on that list. It cannot go off-script.

Tool is the most universal, technically precise term. Everything else below is built on top of this idea.

Skill: a Tool With Personality

A skill is a tool that has been packaged with extra instructions and context.

Imagine the difference between a raw tool and a skill built on the same tool:

A raw tool — minimal description, no guidance:

{
  "name": "search_cases",
  "description": "Search legal cases"
}

A skill — same tool, with instructions and context:

{
  "name": "search_cases",
  "description": "Search legal cases by keyword and location",
  "instructions": "Always filter by jurisdiction first. Summarise results in plain English. Flag anything from the last 6 months.",
  "when_to_use": "When the user asks about finding, looking up, or researching cases"
}

A skill is opinionated. It includes guidance on when to use it, how to use it well, and what to do with the result.

Analogy: A tool is a hammer. A skill is a trained carpenter who knows when to use the hammer, how hard to swing, and what not to hit.

Plugin: the Shopfront Version (Mostly Retired)

A plugin was an early attempt to let third parties publish their own tools in a public AI marketplace.

ChatGPT launched plugins in 2023 — you could install one for OpenTable and the AI could book restaurants, or one for Kayak to search flights. Companies would publish their plugin to an app store, and users could pick which ones to activate.

The idea was straightforward: install a plugin, and the AI gains new abilities. But in practice, plugins were complicated to build, hard to trust, and difficult to use well. OpenAI retired them in 2024.

The concept lives on — just under different names and with better infrastructure. Which brings us to MCP.

MCP: the Universal Plug Socket

MCP stands for Model Context Protocol. It is an open standard created by Anthropic in 2024.

To understand why it matters, consider the problem it solves:

Without MCP:

  App A ──custom code──► AI assistant 1
  App A ──custom code──► AI assistant 2  (rewrite everything)
  App B ──custom code──► AI assistant 1  (rewrite again)
  App B ──custom code──► AI assistant 2  (rewrite again)

  Every combination needs its own integration.

With MCP:

  App A ──MCP──► Any MCP-compatible AI
  App B ──MCP──► Any MCP-compatible AI

  Build once. Works everywhere.

MCP is like the USB standard for AI tools — build your tool once in the MCP format, and any MCP-compatible AI assistant can use it without custom integration code.

Today, tools built with MCP work with Claude, Cursor, and any other app that supports the standard. It is the infrastructure that makes the ecosystem interoperable.

Analogy: Before USB, every device had its own proprietary cable. MCP is USB-C for AI tools.

Hook: the Silent Observer

A hook is completely different from tools and skills — and this is where most people get confused.

A hook is not something the AI uses. It is a behind-the-scenes trigger that your system sets up to watch what the AI is doing.

Here are some examples of things hooks can do — note that the AI never sees any of this:

{
  "before_tool_call": "log_to_audit_trail",
  "after_response": "notify_manager_on_slack",
  "before_email_send": "check_user_has_permission"
}

The AI has no idea hooks exist. They fire invisibly in the background, triggered by events in the agent’s lifecycle.

Analogy: A hook is like a security camera in a store. The customers (the AI) do not interact with it. It just watches and records — or intervenes if something seems wrong.

Hooks are a governance and safety tool. They let developers add oversight, logging, and guardrails without changing what the AI does.

Action: Same Thing, Different Brand

Action is just another word for tool, used by different companies:

OpenAI     →  "GPT Actions"
LangChain  →  "Actions"
Zapier     →  "Actions"
Anthropic  →  "Tools"
Google     →  "Function Calls"

If you see “action” in an AI context, substitute “tool” in your head and you will not go wrong.

Function Calling: the Technical Plumbing

Function calling is the technical name for the mechanism that makes tools work under the hood.

When the AI decides to use a tool, it does not press a button. It outputs a structured message:

{
  "skill": "search_cases",
  "keyword": "landlord disputes",
  "location": "SG"
}

Your application catches this, runs the actual task, and sends the result back. Developers talk about “function calling” when they are building this plumbing. End users never need to hear this term.

How They All Fit Together

flowchart TD
    A([You send a message]) --> B[Agent reads message and skill list]
    B --> C{Needs a skill?}
    C -->|YES| D[Requests a skill via Function Calling]
    C -->|NO| E([Replies directly])
    D --> F[Hook fires — logs the action]
    F --> G[System runs the task]
    G --> H[Result returned to Agent]
    H --> I([Agent writes your final reply])

    J[MCP Server] -->|exposes tools via standard protocol| B

Quick Reference

Term	What it is	Why it matters
Tool	A specific task the AI can request	The foundation — every other concept builds on this
Skill	A tool with added instructions and context	Better instructions lead to better AI behaviour
Plugin	A publicly listed tool — mostly retired	Replaced by MCP and similar standards
MCP	A universal standard for publishing tools	Build once, works with any compatible AI
Hook	A background trigger for oversight and logging	Keeps humans in control without slowing the AI down
Action	Just another word for tool	Knowing this saves confusion when reading different platforms
Function Calling	The technical mechanism behind tool use	The plumbing that makes everything above actually work

The One-Sentence Version

Tools are what the AI can do. Skills are tools done thoughtfully. MCP is the standard that makes tools portable across any AI. Hooks are the oversight layer humans put around everything. Plugins and Actions are just older or platform-specific names for the same ideas.

Once you see the pattern, the jargon stops being confusing — it is all just different layers of the same system.