In March 2023, OpenAI opened the ChatGPT plugin store and the market went wild. Expedia, Zapier, and Wolfram Alpha rushed to publish integrations. Independent developers did too. Within twelve months, more than 1,000 plugins had been published. Twelve months later, OpenAI closed the store. All plugin-based chats died on April 9, 2024.
The cycle repeated. Anthropic released the Model Context Protocol in November 2024, an open standard for connecting AI agents to external tools. The community ran again. SKILL.md files appeared, MCP servers, skills marketplaces. Another wave. And now, in mid-2026, silence again.
The question nobody asks out loud is the right one: what actually happened? And how much is an AI skill worth when it works?
The Pattern We Already Know
The plugins-custom GPTs-MCP cycle is not a technology story. It is a market story. And the AI extensions market followed the same arc as any platform that opened its doors to third parties: initial euphoria, saturation of low-quality supply, brutal consolidation.
In 2023, most plugins did exactly the same thing: search the internet, connect to a public API, summarize something. They were wrappers. And like all wrappers, they lived off the distance between what the model could do alone and what it could do with help. That distance shrank fast.
When the model itself started browsing the web, executing code, and connecting to files, most of the plugin ecosystem lost its reason to exist. The platform ate the builders who had built for the platform.
With MCP something similar happened, but more sophisticated. The protocol is genuinely good: an open standard, donated by Anthropic to the Linux Foundation in December 2025, adopted by OpenAI, Google, Microsoft, and AWS. There are already more than 10,000 active public MCP servers and the SDKs accumulate around 97 million monthly downloads. This is not a dead project. It is infrastructure.
But the builders who flooded the market in 2025 with generic skills faced the same problem as plugins: if your skill does something the base model already knows how to do, you don't have a product. You have a demo.
What Actually Killed the Hype
It was not bad technology. It was a misunderstanding about which part of the value chain they were capturing.
The 2024-2025 wave produced thousands of skills connecting ChatGPT or Claude to public APIs, open databases, and third-party services. Useful for demos. Useful for learning. Hard to convert into something someone pays for or uses sustainably.
The structural problem is this: a generic skill has no proprietary context. And without proprietary context, the model can already do the same thing with a search or with its training. The value of an AI integration is not in the technical connection. It is in the data, the workflow, and the business rules that connection exposes to the model.
When Block implemented MCP agents internally, it reduced its token costs by 98.7% company-wide. Not because it connected the model to a public API. Because it encoded proprietary workflows, real business context, and specific permissions into servers only they control. That is not a marketplace skill. It is a competitive advantage built on the protocol.
Those who disappeared from the market built for the category. Those who remain built for their organization.
When a Skill Is Worth What It Costs to Build
The honest question is not whether skills work. It is what they work for.
There are two types the market separated fairly clearly:
Read skills, those that bring information to the model, are under constant pressure. Current models search, reason, and synthesize increasingly well. Unless the data source is internal and non-public, a generic read skill has a short expiration date.
Action skills, those that execute something in the real world on behalf of the user, are a different story. Creating a ticket, moving a file, updating a record, triggering a workflow. Those the base model does not replace because they require permissions, authentication, and organizational context the model does not have by default. And when well built, the ROI is verifiable: up to 70% reduction in AI operational costs and 300% return in 18 months according to documented enterprise implementations.
The problem is that action skills are harder to build. They require real engineering, governance, identity management. They are not a weekend of code and a README.
The Silence Trap
Nobody talking about AI skills in 2026 does not mean the topic died. It means it matured. And maturity always displaces the early enthusiasts.
41% of software organizations already have MCP servers in production. 28% of Fortune 500 companies run their own MCP servers. That is not silence: it is that the topic stopped being interesting to the media and became operational for companies.
What disappeared was the illusion that anyone could build a generic skill and capture value. What remains is the real work: mapping a specific business flow, encoding it into a governed integration, connecting it to the model with the right permissions, and measuring the result.
That work does not generate headlines. It generates results.
The Point
The skills hype was real. The market emptying was too. But confusing the end of the hype with the end of the topic is the most expensive mistake an organization can make right now.
Generic skills died because they had nothing proprietary to offer. Specific skills, those that encode a team's knowledge, a company's workflows, access to internal data, are today one of the most direct ways to extract real value from AI.
The question is not whether your organization needs AI skills. It is whether the ones it could build have enough proprietary context to justify the effort. That assessment is worth doing well. If you want to analyze it together, reach out to macareno.net.
Sources
- ChatGPT Plugins vs MCP: What Changed and Why It Matters — Unmarkdown
- MCP Adoption Statistics 2026 — Digital Applied
- The Enterprise MCP Guide 2026 — The Agentics
- Are AI Agents Overhyped? An Honest 2026 Reckoning — Taskade