Agent skills teach an AI agent how to do something; MCP connects it to something. A skill is a folder of instructions and optional scripts. MCP, the Model Context Protocol, is a standard way to reach outside tools and data. Use a skill for a repeatable procedure, use MCP for live access to another service, and use both when you need your team's way of working on top of that connection. They solve different problems, so the real question is which gap you have.
Key facts
- Skill: know-how in a folder: a SKILL.md file, plus optional scripts and templates
- MCP: a protocol that connects an agent to outside services through a server
- Choose a skill when: you keep pasting the same steps, rules or house style into chat
- Choose MCP when: the agent needs live data or actions in another service
- Start-up cost: about 100 tokens per skill; MCP depends on the agent
- Together: MCP reaches the system, and a skill says how your team uses it
On this page
- Agent skills vs MCP compared
- What skills are good at
- What MCP is good at
- Which should you use? A quick test
- How skills and MCP work together
- Skills vs MCP on token cost
- Skills vs MCP on safety
- Frequently asked questions
Agent skills vs MCP compared
Anthropic created both, and its Claude Code documentation sums up the split: "MCP connects Claude to external services. Skills extend what Claude knows, including how to use those services effectively." Here is how they differ in practice:
| Compared | Agent skill | MCP |
|---|---|---|
| What it is | A folder with a SKILL.md file | A protocol, used through an MCP server |
| What it adds | Knowledge, procedures, templates and scripts | Tools and data in other systems |
| Who uses it | The model reads it and follows it | The model calls it, and the server does the work |
| What you set up | Copy a folder or run one install command | Add a server, often with a login or API key |
| Runs where | On the same machine as your agent | Locally or on a remote server |
| Example | A code review checklist or a deploy procedure | Query a database or post to Slack |
| Good at | Consistency | Reach |
| Main risk | Instructions and scripts you did not write | A new connection with its own permissions |
| Standard | Agent Skills, published December 2025 | MCP, created in late 2024 and now at the Linux Foundation |
New to either? Start with what are agent skills? or what is MCP?.
What skills are good at
Skills are best when the agent already has the tools and lacks the method. Writing minutes, reviewing code, following a brand guide and preparing a report are all jobs where the steps matter more than the connection.
A simple test: if you have typed the same instructions into a chat three times, that is a skill. It needs no server, no login and no API key, and it works offline if your agent does. It can also move between agents, because the same folder works in Claude Code, Codex, Cursor and others.
The limit is reach. A skill cannot log in to your accounts system by being well written. It can include a script that calls a service, but then you are managing that script and its keys yourself.
What MCP is good at
MCP is best when the agent needs something outside its own session: today's figures from a database, a message sent to Slack, a page opened in a browser. The server holds the connection and handles authentication, so the agent gets a clean tool to call.
It is the better choice when several people or apps need the same governed connection. One server can serve many clients, and you change it in one place. See how to add MCP servers to Claude or, for free local models, LM Studio plugins and MCP.
The limit is judgement. An MCP server tells the agent what it can do, and not how your team wants it done.
Which should you use? A quick test
| Your situation | Use |
|---|---|
| You keep pasting the same instructions into chat | Skill |
| You want a consistent format, tone or checklist | Skill |
| The agent keeps making the same mistake with your project | Skill |
| The agent needs live data from another app | MCP |
| The agent must act in another app, such as sending or updating | MCP |
| Several people need one managed connection, with its own login | MCP |
| The agent can reach a tool but uses it badly | Skill on top of MCP |
| You want to share a whole set-up with a team | A plugin containing both |
| The rule must apply in every session, with no exceptions | Neither: use custom instructions or a hook |
If a command-line tool already does the job, you may not need an MCP server at all. A skill can tell the agent how to use that tool, and some of the skills in our weekly chart work this way.
How skills and MCP work together
Many set-ups use both, each for its own job. Anthropic gives the example of a skill that holds your team's database schema and query patterns, used alongside a database MCP server. Another is a /post-to-slack skill holding your message format, on top of a Slack connection.
Here is a small business version. Say your accounts software has an MCP server:
- MCP lets the agent read overdue invoices.
- A skill holds your chasing rules: wait seven days, use this tone, never chase a disputed invoice, and leave every email as a draft for a person to approve.
- The result is the same process each time, rather than a fresh interpretation each time. Our invoice-chasing agent guide shows a full build.
Plugins package this up. In Claude Code a plugin can hold skills, hooks, subagents and MCP servers, and in Codex a plugin can hold skills and connectors, so a team installs one thing.
Skills vs MCP on token cost
The common claim that "skills are cheap and MCP is expensive" is out of date. Both now try to load as little as possible, and the answer depends on your agent.
- Skills: about 100 tokens per skill at start-up for the name and description. Full instructions load on demand, and the specification recommends keeping them under 5,000 tokens.
- MCP in Claude Code: tool names load at the start and full details load on demand, because tool search is on by default. Anthropic's documentation rates this as low cost until a tool is used.
- MCP elsewhere: this varies. An agent that loads every tool definition up front can spend a lot before you start. In November 2025 Anthropic described one example of this, a task that fell from 150,000 tokens to 2,000 once tools were loaded only as needed.
This matters most on local models, where memory is short. See agent skills on local models for the sums. In Claude Code, /context all shows what each tool is using.
Skills vs MCP on safety
Both give an agent more power, in different ways, and neither is safe by default.
- Skills bring in instructions, and sometimes scripts, that run with your agent's access to your files and keys. Fake skills have already been used to steal passwords. Run our five-point check first.
- MCP servers add a connection, and each connection is a permission. Give each one the narrowest access that works, as what is MCP? explains.
- Both: keep a person in front of anything you cannot undo, such as sending, paying or deleting.
For UK businesses, the data route matters. A skill that only holds instructions moves no data on its own, but an MCP connection to customer or staff records can. If yours does, read AI agents and UK GDPR before you switch it on. This is general information, not legal advice.
Frequently asked questions
What is the difference between agent skills and MCP?
Agent skills give an AI agent knowledge and procedures, stored as a folder with a SKILL.md file. MCP, the Model Context Protocol, connects an agent to outside tools and data through a server. A skill teaches the agent how to do something; MCP gives it something new to reach.
Do agent skills replace MCP?
No. They solve different problems. A skill can tell an agent how to use a service, but it cannot by itself provide a logged-in connection to it. MCP handles that connection and its authentication. Some jobs that once used an MCP server can now use a skill with a command-line tool, but the two remain separate.
Can a skill use an MCP server?
Yes, and this is often the best set-up. The MCP server connects the agent to a system such as Slack or a database, and the skill sets out how your team uses it, such as your message format or common queries. Plugins in Claude Code and Codex can bundle both.
Which uses more tokens, skills or MCP?
It depends on the agent. A skill costs about 100 tokens at start-up, with full instructions loaded on demand. An MCP server used to load every tool definition up front, which could be expensive, but Claude Code now loads tool names first and fetches full details when needed. Check how your own agent handles it.
Are skills safer than MCP?
Neither is safe by default. A skill brings in instructions and possibly scripts that run with your agent's access. An MCP server adds a new connection, and each connection is a permission. Check who made either one, give it the narrowest access that works and keep a person in front of anything you cannot undo.
Should a small business use skills or MCP first?
Start with a skill if your problem is consistency, such as house style or a checklist, because it needs no connection to anything. Add MCP when the agent needs live data from another service. If that data includes customers or staff, check your UK GDPR position first.
Read next
- What are agent skills?
- What is MCP?
- Agent skills on local models
- Are agent skills safe to install?
- Claude MCP: how to add servers
- Top 20 agent skills this week
Sources
- Extend Claude Code, Anthropic docs - checked 6 November 2026
- Connect Claude Code to tools via MCP, Anthropic docs - checked 6 November 2026
- Code execution with MCP, Anthropic - checked 6 November 2026
- Agent Skills specification - checked 6 November 2026
- What is the Model Context Protocol? - checked 6 November 2026
- Build skills, OpenAI docs - checked 6 November 2026
