Every new wave of technology brings a familiar chorus: "This will kill everything that came before it." GraphQL was going to kill REST. MCP was going to kill all APIs. Now MCP apps are supposedly going to kill all web UIs. On the latest episode of the Naftiko Capabilities podcast, host I sat down with Kevin Swiber — a longtime collaborator and deep thinker on APIs and AI integration — to cut through the noise and talk about what's actually happening with MCP, agent skills, and the future of AI-powered workflows.

The Hype Problem

We both acknowledge a frustrating reality: the technology is moving faster than most people can absorb it. And the sensationalism isn't helping. Kevin shared that after interviewing roughly 30 enterprise professionals last month, he found the majority are still using AI as a glorified Stack Overflow — copying code into ChatGPT, getting a response, and pasting it back into their IDE. Despite the billion-dollar revenue figures being thrown around, most practitioners aren't deep in the rabbit hole of agents and protocols.

That disconnect matters. "I think it is actually very important that we are responsible in how we talk about these things because they actually do have real impact in the world," Kevin said. Telling people to go all-in on a single technology can box them into career dead ends.

MCP: Powerful, But Not a Silver Bullet

Model Context Protocol has become the buzzword of the moment, and for good reason — it gives AI agents a standardized way to access tools like Salesforce, Notion, GitHub, and more. But Kevin is quick to point out that MCP's real sweet spot is narrower than the hype suggests.

MCP shines when a human user is involved and self-selecting integrations — choosing which accounts and tools to wire into their AI experience. But for headless agents running in the background where you already control the code and know all the possible actions upfront? "MCP is certainly overkill," Kevin said. "I would say you don't need it at all."

There's also the context window problem. Many MCP hosts dump the entire tools list into the context, quickly eating up tokens with information that provides no value. The community has been working on solutions — semantic search on the host side, Anthropic's new tool search tool — but it's still a pain point.

Enter Agent Skills

This is where agent skills enter the picture. Primarily driven by Anthropic and Block, the agent skills spec addresses something MCP doesn't: not just what tools are available, but how to use them effectively.

Kevin offered an analogy from Anthropic: imagine walking through a hardware store. Your MCP server is the aisle you're browsing — plumbing, electrical, whatever. The agent skill is the store employee who tells you exactly what to buy and how to fix your leaky faucet. Having tools is one thing. Knowing the workflow to apply them is another.

Agent skills also got a key design pattern right early: progressive disclosure. Rather than dumping all available context into the window at once, skills surface metadata first and let the LLM dig deeper only when needed. MCP is catching up on this front, but skills had the head start.

Skills Are for Everyone

Perhaps the most compelling argument for agent skills is accessibility. MCP servers require engineering effort — you're writing code, managing infrastructure, and configuring integrations. Skills, on the other hand, are essentially text files. A marketing professional can craft a skill based on their own workflow without writing a single line of code, share it with colleagues, and start seeing immediate benefit.

"Really anyone in your marketing department can start crafting a skill based on their own workflow to try to automate their workflow," Kevin explained. As people build and share skills, common patterns emerge organically, leading to better organizational knowledge about how work actually gets done.

The Takeaway

The conversation between Kevin and I counters the all-or-nothing narratives dominating the AI tooling space. The real answer, as usual, is nuanced: MCP and agent skills serve different purposes and work best together. MCP provides the what — access to tools and systems. Agent skills provide the how — the knowledge and workflows that make those tools useful.

Instead of declaring winners and losers, the industry would benefit from understanding when each approach is the right fit and building accordingly. As Kevin put it: "Can we just enjoy something for a minute?"

‍