MCP vs Skills — 3-Line Answer and Decision Guide

Answered in 3 lines

1. MCP = a connection to external systems (APIs, DBs, CLIs, filesystems, etc.)
2. Skills = knowledge / playbook kept inside the agent (rules, templates, best practices)
3. Use both when you need both — they are not exclusive. Need a connection? MCP. Want to teach knowledge? Skills.

At-a-glance mapping

What you want to do	MCP	Skills
Call an external API	✅	❌
Query an internal database	✅	❌
Access the filesystem	✅	❌
Teach project coding conventions	❌	✅
Teach "how to create a PR" procedure	❌	✅
Provide domain expertise (laws, jargon)	❌	✅
Provide Markdown templates	❌	✅
Teach how to use a CLI (gh, aws, etc.)	❌	✅

Common search questions, answered in 3 lines

Q: Should I build MCP or Skills first?

A: Skills first. Skills start from a single Markdown file and pay off immediately. MCP carries higher implementation and operational cost — only build one when you are certain you need an external connection.

Q: What is skill.md?

A: A Markdown file that teaches an AI agent domain knowledge. Place it as SKILL.md and the agent loads it on demand. It is plain Markdown with frontmatter (name, description) and a body (instructions / procedures).

Q: Isn't MCP more powerful? Isn't Skills just a watered-down MCP?

A: No. MCP and Skills serve different responsibilities. MCP supplies "what you can access"; Skills supplies "what you know and how you decide." When a CLI already exists, gh CLI + Skills is often more token-efficient than building an MCP.

Q: Are Cline / Cursor / Vercel Skills the same?

A: They share the same specification base (Agent Skills Specification). The SKILL.md format is identical, but placement paths differ: Claude Code uses .claude/skills/, Cursor uses .cursor/rules/, Cline uses .pi/skills/, Vercel manages them via npx skills. See What is Skills for details.

Q: I heard MCP is heavy (high token consumption)

A: True. Each MCP server consumes context just for its tool definitions. Connecting 10 MCPs can cost tens of thousands of tokens. Mitigations: (1) disconnect unused MCPs, (2) use Tool Search / Deferred Loading, (3) replace MCPs with Skills + CLI where possible. The structural reason is explained in understanding-llm / MCP Context Cost.

Q: Is this different from sub-agents?

A: Yes. Skills are knowledge; sub-agents are separate processes. Skills are "an instruction sheet for how to do something"; sub-agents are "specialist staff." A sub-agent can carry its own Skills and MCPs. See Sub-agents for details.

Q: What is the typical pattern when using both?

A: "MCP for connection, Skills for behavior" is the dominant pattern. Examples:

• MCP github-mcp fetches PR data → Skills pr-review apply review criteria
• MCP postgres-mcp reads the DB → Skills db-conventions enforce naming rules
• MCP slack-mcp posts messages → Skills notification-style unify tone

Decision flow (decide in 10 seconds)

Going deeper

What you want to know	Page
Detailed MCP vs Skills selection	MCP vs Skills (full version)
Skills structure and format	What is Skills
MCP structure and protocol	What is MCP
How to create a Skill	How to Create Skills
How to build an MCP	MCP Development
Why Skills must be a separate layer	understanding-llm / Part 5
Why MCP becomes a context cost	understanding-llm / Part 6