Two acronyms, a year apart, get conflated constantly: MCP and Skills. Both came out of Anthropic. Both get pitched as "how you make agents actually useful." And they solve two completely different problems — one is about reach, the other is about judgment.
What MCP Actually Solves
Anthropic introduced the Model Context Protocol in November 2024 to fix a specific problem: models were "constrained by their isolation from data—trapped behind information silos", and every new data source needed its own bespoke integration. MCP is an open standard for a universal, two-way connection between an AI assistant and an external system — Slack, GitHub, Postgres, Google Drive. One protocol instead of N × M custom wires.
The architecture is client-server: an MCP server exposes a system's tools, resources, and prompts; an MCP client — Claude, or any AI app — consumes them. Anthropic shipped pre-built servers for the obvious targets at launch. The question MCP answers is simple: can the agent reach this system at all?
What Skills Actually Solves
Almost a year later, Anthropic introduced Agent Skills (October 2025) to fix a different problem entirely: even with full access to a system, an agent doesn't know your way of using it. Their own framing: building a skill is like putting together an onboarding guide for a new hire.
A skill is a folder with a SKILL.md file at minimum — name and description in YAML frontmatter, loaded into every system prompt so Claude knows when to reach for it. That's level one of what Anthropic calls progressive disclosure. If the task matches, Claude reads the full SKILL.md body — level two. If that references bundled files (a forms.md, a Python script), Claude reads or runs those only when it actually needs them — level three and beyond. The skill can be arbitrarily large because almost none of it sits in context by default. The question Skills answers: does the agent know your specific process here?
MCP: can it reach the system?
A standard connection. Claude can now see and act on a system it couldn't reach before.
Skills: does it know what to do?
Packaged expertise. Claude does the task the way your team actually does it.
Most real work needs both
Anthropic's Own Answer
Anthropic doesn't frame these as competing. In the Skills launch post, they say directly that they plan to explore how Skills can complement MCP servers — teaching agents the more complex workflows that involve the external tools MCP connects them to. MCP is the access; Skills is what the agent does once it has that access. Neither one substitutes for the other, and most non-trivial setups end up needing both — an MCP server to reach the system, a skill to encode the runbook for using it the way your team actually works.
IBM's Enterprise Lens
IBM's MCP explainer is more architecture-pedagogy than Anthropic's original announcement, and it adds a line worth sitting with: MCP is "not an agent framework, but a standardized integration layer for agents accessing tools". It names LangChain, LangGraph, and crewAI explicitly as frameworks MCP complements rather than replaces — MCP doesn't decide when a tool gets called, the model does.
The more useful contribution is the risk section. IBM flags that MCP integrations are brittle by construction — every provider's API schema is different, and "the slightest change in these tools can result in the collapse of this entire AI workflow infrastructure." It's also candid about what community MCP servers don't give you out of the box: built-in identity, versioning, or audit trails. That's the actual reason enterprise deployments end up wrapping MCP in a gateway — auth, rate-limiting, a server registry — rather than pointing agents at raw community servers directly.
The Two-Question Test
When a task feels like it needs "more agent capability," the useful move isn't reaching for whichever acronym is trending. It's asking, in order:
- Can the agent reach the system at all? If no — it has no connection to your ticketing tool, your database, your internal API — that's an MCP problem. Build or connect a server.
- Does the agent know your specific process once it's there? If no — it can query the database but doesn't know your escalation thresholds, your report format, which channel to notify — that's a Skill problem. Write the runbook down.
If the answer to both is no, you need both, in that order — access before methodology. If the answer to both is yes, you don't need either. You need to find out why the agent still isn't doing the thing.
Where It Breaks
Two failure modes worth knowing before you build either. Skills carry a security cost Anthropic is explicit about: a skill is instructions and code, which means a malicious one can direct an agent to exfiltrate data or take unintended actions just as easily as a legitimate one can save you time. Install from trusted sources, audit anything you don't control the source of.
MCP carries the brittleness IBM flags — a schema change on the far end of a server can quietly break every skill and workflow built on top of it, with no local warning until something fails downstream. Neither problem is a reason to avoid either technology. They're the reason "which one do I need" is worth asking on purpose, instead of wiring up whichever one you read about most recently.