Claude Skills: What are Claude Skills and how are they different than MCP?

If you're in the techie AI world, you’ve probably heard of the Model Context Protocol (MCP) and maybe even experimented with it. And it’d be easy to conflate it with Anthropic's new “Skills”. But once you separate them in your head, you’ll see they serve very different roles, and when you put them together, they have some pretty powerful capabilities.

At its heart, MCP is about access. It defines a standard protocol for AI models (or agents) to reach out to external systems: your database, API, file system, messaging system, whatever sits outside the model’s native knowledge. In effect, you build or adopt an MCP server that exposes certain tools or data sources under a common interface. The model becomes capable not just of chat-style responses, but of interacting with your systems in a structured way.

By contrast, “Skills” are about process and behavior. They don’t focus on the underlying connection to external systems (that’s what the protocol is for). Instead, they define how the system should handle a given task. They package instructions, scripts, and resources for the Claude model to load when faced with a particular workflow.

For example: define the tone, decision-branches, escalation criteria, template formats, etc., all the stuff that makes a workflow repeatable and consistent.

The practical distinction is that MCP says “what the model can access.” Skills say “how the model should act once it has access (or just via instructions)”.

Imagine you’re creating a support-agent workflow. First you build an MCP connection that gives the model access to your ticketing system, customer database, maybe even email logs. That is your plumbing. Then you define a “Support Skill” that says: check customer history, search knowledge base via the MCP, if you don’t find a match, open a ticket with the correct template (again, MCP), maintain a friendly yet professional tone, escalate if there’s mention of a refund over $500, etc. That is your behavior. You can see that they're not mutually exclusive. They're sort of different layers of the stack.

You’ll see scenarios where you might pick one over the other. If you simply want the model to access a data source or an API-- “go fetch repo status” or “query the CRM”-- then implementing only an MCP server might suffice. On the other hand, if you have a strictly defined task that doesn’t involve external systems-- for instance, “take this daily stand-up text and produce a summary with blockers and next-steps”-- you might implement it purely as a Skill (instructions and a template) without any external tool access.

By combining them, you give the model more capabilities and a consistency in how it uses them.

It’s worth noting the ecosystem factors. MCP is an open protocol introduced by Anthropic to standardize how models integrate with external systems. On the other hand, Skills (as currently packaged by Anthropic) is a more proprietary layer that I imagine OpenAI will quickly match.

If you’re getting started, here’s a simple path to using Skills: Identify one workflow that feels like it should be predictable and repeatable. Write out the steps, rules, tone, escalation points and make this your Skill. Next, decide if this workflow requires external system access for data lookup, automation, ticketing, etc. If needed, set up an MCP server or, hopefully, re-use an existing one. Then hook up the Skill to the server.

MCP is basically about “what the AI can reach”. Skills are about “how the AI should behave”.