Commands vs MCP vs Skills

AI coding agents like Claude Code can be extended in three main ways: commands, MCP tools, and skills. Each serves a different purpose and has different trade-offs.

Viktor Farcic explains this better than I could, so watch the video first, then use the summary below as a quick reference.

Quick Reference

Commands (Slash Commands)

Pre-defined prompt templates stored as markdown files, triggered explicitly by the user (e.g., /commit, /review).

• Simple to create and maintain (just markdown files)
• Consistent execution every time
• Must be invoked manually; the LLM won’t discover them on its own
• No standard location across agents (Claude Code uses .claude/commands/, Cursor uses .cursor/rules/)

MCP Tools

Server-based tools exposing external capabilities through a standardized protocol. The LLM decides when to call them based on context.

• Extends what the agent can do (query APIs, databases, monitoring systems)
• Standardized interface that works across all MCP-compatible agents
• Invoked automatically; the LLM picks the right tool for the job
• Higher implementation cost, network latency, and security considerations

For more on MCP, see What is MCP?.

Skills

Markdown-based workflows that the LLM can invoke both explicitly and automatically when the context matches. Only skill descriptions load into the system context initially; the full instructions inject when triggered.

• Dual invocation: explicit (/skill-name) or automatic based on context
• Minimal context usage (descriptions only until triggered)
• Teaches the agent how to use its abilities well
• Distribution across projects is still a challenge

When to Use What

The key insight: MCP gives agents abilities. Skills teach agents how to use those abilities well. Commands are for when you want full control over when something runs.

Source

• Full transcript on DevOps Toolkit