Claude skills vs MCP: what each does, when to use which

Searches for “Claude skills vs MCP” usually come from someone about to build one of the two. The short answer is that they solve different problems: an MCP server is a connection, a skill is know-how. This page sets them side by side using Anthropic’s Claude Code docs and the MCP project’s own pages, then shows how they work together, with Lookbk’s own MCP server and skills as a worked example. For prompts, commands, plugins, hooks and CLAUDE.md, see Claude Code skills vs prompts, commands and plugins.

What each one is

A skill

A folder with a SKILL.md file: a header with a name and a description, then instructions. It can carry scripts, reference files and templates beside it. Claude Code keeps a list of every skill’s name and description in context and loads a skill’s body only when it is used. Claude Code’s skills follow the Agent Skills (opens in a new tab) open standard, so the same folder also works in other agents that read it. What are agent skills? covers the idea in full.

An MCP server

The MCP project calls the protocol “an open-source standard for connecting AI applications to external systems”, and compares it to a USB-C port. An MCP server is a program that offers three kinds of thing to the AI application connected to it:

• Tools: functions the model can call to act, such as an API call or a database query.
• Resources: data that gives context, such as file contents or database records. In Claude Code you reference them with @.
• Prompts: reusable templates. Claude Code lists each one as a command, /servername:promptname (MCP).

A server runs either on your machine, talking over standard input and output (stdio), or remotely over Streamable HTTP. Claude Code’s docs say to connect one “when you find yourself copying data into chat from another tool”, so Claude can read and act on that system directly.

Claude skills vs MCP, side by side

When to use a skill, and when MCP

A useful test: if you could do the job by pasting instructions into the chat, it is a skill. If you would have to paste data out of another tool, it is MCP.

How skills and MCP work together

Anthropic’s advice is to “use both together: MCP for connectivity, Skills for procedural knowledge.” In practice a skill’s steps call MCP tools by name. Claude Code gives every MCP tool a full name, mcp__<server>__<tool>, and its docs say to use that name in a skill’s allowed-tools list, which lets the skill call those tools without a permission prompt during the turn it runs.

---
name: pick-a-visual-skill
description: Find and install a Lookbk agent skill for a visual job, such as a launch video, a pitch deck or an OG image. Use when someone asks which skill to use for making a video, slides, a social image or a landing page.
allowed-tools: mcp__lookbk__search_skills mcp__lookbk__get_skill mcp__lookbk__get_install_command
---

1. Call `search_skills` with what the person wants to make, in their words.
   Show the top three with whether each is free.
2. When they pick one, call `get_skill` and tell them what it needs
   (Node.js, Chrome, FFmpeg) before going further.
3. Call `get_install_command` with agent "claude-code" and show the command.
   Run it only after they say yes.

The server supplies the catalog; the skill supplies the judgment about what to show, what to check and when to stop and ask. Neither alone would do the job as well.

Two more places they meet. A Claude Code plugin can ship both at once: skills under skills/ and server definitions in .mcp.json. And an MCP server author can write server instructions, which Claude Code’s docs say help Claude decide when to look for the server’s tools, “similar to how skills work”.

A real example of both: Lookbk

Lookbk has both, and the split shows the difference well.

• The skills are SKILL.md folders that make things on your machine. Editorial OG Card, for instance, holds scripts that read your posts into cards.json and render 1200×630 PNGs with the Chrome already installed, plus the card template and two fictional examples.
• The Lookbk MCP server, at https://lookbk.dev/api/mcp, gives an agent four read-only tools over the catalog: search_skills, get_skill, list_video_skills and get_install_command. It offers no resources or prompts, keeps no session, and never installs anything. get_install_command returns the command as text, and your agent runs it in your terminal like any other command.

claude mcp add --transport http lookbk https://lookbk.dev/api/mcp

So the MCP server answers “which skill, and how do I get it?”, and the skill answers “how is this card made?”. The server renders nothing, because rendering needs your files and your browser, so that work lives in the skill. And a skill could only search a catalog that keeps changing if a copy of the catalog were packed into it, which is what the server avoids.

Free

image

Editorial OG CardClaude CodeCodexCursor+3 more agents

Free

video

Kinetic Title DropClaude CodeCodexCursor+3 more agents

Free

image

Blueprint Architecture DiagramClaude CodeCodexCursor+3 more agents

Free

slides

Swiss Grid DeckClaude CodeCodexCursor+3 more agents

Context cost and trust

• Skills: Claude Code gives the list of skill names and descriptions 1% of the context window, and drops descriptions of the least-used skills first when it overflows. A skill’s body costs nothing until it is used.
• MCP: with tool search on, which is the default on recent models, Claude Code defers tool definitions until Claude needs them, so adding servers has little effect on context.
• Trust: Claude Code’s docs say to verify you trust each MCP server before connecting it, since servers that fetch outside content can expose you to prompt injection. A skill’s scripts run on your machine, so read them first; Safety checks before running a skill lists what to look for.

Claude skills vs subagents

A subagent is a third thing: a separate assistant with its own context window, system prompt, tool access and permissions, which does a side task and returns a summary. Claude Code’s subagent docs say to consider a skill instead “when you want reusable prompts or workflows that run in the main conversation context rather than isolated subagent context.”

All three combine. A subagent’s skills field injects the full content of the listed skills when it starts. Its mcpServers field gives it MCP servers the main conversation doesn’t have, so their tool descriptions stay out of your main context. A skill can also run in a forked subagent with context: fork. The fuller comparison, with commands, plugins, hooks and CLAUDE.md, is in Claude Code skills vs prompts, commands and plugins.