MCP Started As A Tool-Use Protocol. It Became The Publishing Protocol When Nobody Was Looking.

Anthropic shipped MCP (the Model Context Protocol) as a way for Claude to call tools. That was the framing.

It is not how it is being used in 2026.

In the last six months MCP became the publishing layer. HTML Pub ships an MCP server so you can publish from Claude directly. Cursor ships MCP. Windsurf is wiring it in. Every IDE adjacent AI tool is shipping or planning an MCP integration, and the integration that ships first is almost always "publish this thing somewhere real."

That is the second-order shift category leaders are missing.

What MCP Actually Is

In short, it is a spec. Anthropic published it, but the spec is open. Any AI engine can speak it. It is the USB-C of AI tool-use. Standardized. Plugged in once, reusable across every engine.

The publish button your AI was missing turned out to be an MCP server. Connect once, publish from any AI conversation. Not a Claude or ChatGPT only thing. The endpoint is engine-agnostic by architecture.

Why Category Leaders Didn't Ship It

Squarespace has an API.Squarespace has an API. Wix has an API. Webflow has an API. None of them speak MCP.

They speak proprietary REST, with rate limits, with custom auth, with documentation written for human developers reading PDFs at 2am. They were built for a world where humans called APIs and AI was nowhere near the input.

Now AI is the input. The human is describing what they want. The AI is calling the API. And the API needs to be a thing the AI can discover, parse, and use without a human writing custom glue.

MCP solves that, while REST APIs do not.

This is the gap. Category leaders are sitting on closed pipelines built for the world that ended last year. The AI-native challengers shipped MCP servers and skipped the integration layer entirely.

What This Means If You Are Building With AI

Your tooling is converging on MCP as the protocol. Your publishing tool needs to speak it too. Otherwise you are stuck pasting HTML or worse, opening a dashboard.

Three things to look for in a publishing tool that takes MCP seriously:

Native MCP server. Not a Zapier middleware bridge. Not "we have an API and you can hand wire it." Actual MCP server, discoverable, callable from any engine that speaks the protocol.
Engine-agnostic. Works with Claude, ChatGPT, Cursor, Windsurf, whatever you use today. No lock-in to one engine.
Real publish action. Not "create draft" or "save to staging." A literal publish endpoint that returns a live URL.

The tools that get this right become invisible infrastructure. The tools that don't are dashboards your AI cannot reach.

What This Means For Category Leaders

You have months, not years, to ship MCP. The AI native side of the market is converging fast. By the time Squarespace ships an MCP server (that is if they do), the workflow will already be locked into the tools that shipped one in 2026.

This is the protocol question every platform team should be answering this quarter. Not whether to ship an MCP integration, but rather when.

Why HTML Pub Built One

HTML Pub MCP is the publish button your AI was missing. You connect it once. From then on, every AI conversation has a publish endpoint. Describe a page in Claude, get a live URL. Describe a page in ChatGPT, get a live URL. Describe a page in Cursor, get a live URL.

Carrd is an editor, while HTML Pub is a description. They design, we actually build. The MCP server is the difference between describing a page and shipping a page.

This is the architecture the next five years of publishing run on. Quietly. Without anyone announcing it as a category.

If you are building anything where AI publishes something, MCP is your protocol now.

About the author

Yvonne Chow is the head of marketing at Leadpages. She runs HTML Pub on the marketing side and ships pages every week.