Source: Anthropic Blog Extending Claude Skills Mcp 2025 12 19 (Anthropic blog, Dec 19 2025 — https://claude.com/blog/extending-claude-capabilities-with-skills-mcp-servers)

Anthropic’s framing of how Skills and MCP servers compose into effective agents. Not competitors — they solve orthogonal problems. Skills teach Claude how to do something (methodology, workflows, standards). MCP gives Claude access to systems (data, APIs, platforms). The combination is where real leverage comes from.

Key Takeaways

  • Hardware store analogy. “MCP is like having access to the aisles. Skills…are like an employee’s expertise.” An agent without skills is a customer wandering the aisles. An agent without MCP is an expert with no tools. You want both.
  • Skills = methodology, MCP = access. Skills provide procedural knowledge and “how” patterns. MCP provides connectivity and “access to” external data and platforms.
  • Loading models differ. Skills are on-demand (loaded when triggered). MCP is always-available (connected throughout the session). This affects context budgeting.
  • Real-world composition examples from the blog:
    • Comparable company analysis skill orchestrates multiple MCP connections (S&P Capital IQ, Daloopa, Morningstar) to run valuations automatically
    • Notion Meeting Intelligence skill coordinates with Notion MCP to pull prior context and generate structured meeting prep documents
  • The composition pattern is the point. Single-skill and single-MCP usage exist, but the real agent leverage emerges when a skill orchestrates multiple MCPs toward a goal — neither primitive accomplishes that alone.

Skills vs MCP — the side-by-side

AspectSkillsMCP
FunctionTeaches methodologyEnables access
Best forWorkflows, standards, proceduresData access, API calls, external systems
LoadingOn-demand (when triggered)Always available in the session
ContentFiles (SKILL.md, scripts, references)Server processes exposing tools/resources
AuthorsNon-engineers can writeTypically engineer-built

Composition patterns seen in production

  1. Skill-wraps-MCPs pattern. A domain skill encodes the workflow; MCPs provide the data feeds. Example: a pitch-deck skill consults market-data MCPs and brand-asset MCPs, producing a deck without explicit user orchestration.
  2. MCP-triggers-skill pattern. An MCP event (e.g., a ticket update) hands off to a skill for standardized response. Example: a support-ticket MCP event triggers a triage skill.
  3. Multi-skill-coordinated pattern. Multiple skills run in sequence, each calling the MCPs it needs. Example: research skill → analysis skill → presentation skill, each drawing on different MCPs.

Getting started

Relation to other Claude extensibility layers

Open Questions

  • Context budget at scale. When dozens of MCPs are connected and many skills are installed, how does Claude prioritize which skills/MCPs to surface? Blog doesn’t quantify.
  • Skill-MCP authoring symmetry. Skills can be authored by non-engineers; MCPs typically can’t. Does that create a gap where teams can write workflow knowledge but can’t expose the data sources those workflows need?
  • Failure modes. What happens when a skill expects an MCP that isn’t connected? Is there a declarative dependency mechanism or just runtime error?

Try It

  1. Pick one composition. Take a workflow you run manually today (weekly report, customer research, pitch prep). Identify: what methodology does it follow (skill candidate) and what systems does it read from (MCP candidates).
  2. Install one Anthropic skill from https://github.com/anthropics/skills that matches a real workflow. Run it. Notice whether it calls any MCPs you have connected.
  3. Read the follow-on blog Building Agents with Skills for the progressive-disclosure architecture and skill-type taxonomy (Foundational / Partner / Enterprise).
  4. Audit your MCP connections against Essential MCP Servers for 2026. Skills only help if the MCPs they need are reachable.