What is the difference between Agent Skills and MCP?
Agent skills are static filesystem-based instruction packages — Markdown files with YAML frontmatter that teach an AI coding agent how to perform a specific task. They are loaded at startup from a well-known directory, require zero infrastructure, and work entirely offline. MCP (Model Context Protocol) is a live client-server protocol where the AI agent connects to a running MCP server at runtime to access dynamic tools, live data, and external APIs. The key difference: agent skills are instructions baked into the agent's context window; MCP tools are external capabilities the agent calls on demand.
Agent Skills: pros and cons
Pros: no server to run, zero latency, works offline, version-controlled alongside code, portable across all compliant platforms, minimal context overhead via progressive loading. Cons: static only — cannot fetch live data, call APIs, or execute code unless the agent itself does it. Best for encoding expert knowledge, coding conventions, workflow patterns, and domain rules that don't change per-request.
MCP: pros and cons
Pros: can provide live data (weather, prices, database queries), execute code in isolated environments, expose complex multi-step tools, and integrate with any external service. Cons: requires a running server process, introduces network latency, more complex to set up and maintain, not portable without infrastructure. Best for dynamic integrations: querying your database, calling internal APIs, running code sandboxes, or fetching real-time information.
When to use Agent Skills vs MCP
Use agent skills when you want to encode reusable expertise: coding standards, security audit checklists, testing workflows, documentation templates, or framework-specific patterns. Use MCP when you need live data or external actions: querying a Postgres database, calling the GitHub API, running a Python sandbox, or fetching current stock prices. Many teams use both together: agent skills handle the 'how to think' layer while MCP handles the 'how to act' layer.
Can Agent Skills and MCP work together?
Yes. An agent can have both agent skills loaded in its context and MCP servers connected simultaneously. A common pattern: a 'Stripe Integration' agent skill encodes best practices and error-handling patterns (static instructions), while a Stripe MCP server provides live tools to actually call the Stripe API (dynamic actions). The skill teaches the agent how to use the tool correctly; the MCP server provides the tool itself.