On MCP | Clarisix

← All posts<br>TL;DR<br>MCP is a real, useful protocol — but the dashboard-is-dead, AI-replaces-everything narrative depends on something most products don't have: a coherent data foundation. Without it, MCP just makes hallucinations faster. The companies that win the AI-native future are the ones doing the unglamorous work of rebuilding their data foundation now, while everyone else demos chatbots.

Why the hype is half right

Every conversation about software in 2026 eventually arrives at three letters: MCP.

If you spend any time near AI infrastructure, AI tooling, or the people who fund both, you have heard the claim. Model Context Protocol is going to change everything. It is the USB-C of AI. It is the death of dashboards. It is the moment when language models finally stop being chatbots and start being agents that actually do things.

Half of that is true. The other half is selling something.

What MCP actually is

MCP is a protocol. That is the whole story.

It is an open standard, introduced by Anthropic in late 2024, that lets an AI model talk to external systems in a consistent way. Instead of every developer building a custom integration to plug their software into Claude or ChatGPT or any other model, the model speaks MCP, the application speaks MCP, and the two find each other.

Think of it as a translation layer. Before MCP, every AI integration was a bespoke project. After MCP, the integration becomes plug-and-play. A model asks a question. The application responds with structured context. The model makes a decision or surfaces an answer.

This is genuinely useful. I am not going to argue against the protocol itself. Standards are how software ecosystems mature, and AI needed one badly.

But MCP is the transport layer. It is the wire. It is not the truth.

Why everyone is suddenly excited

The excitement has nothing to do with the protocol itself and everything to do with what it suggests is possible.

If a model can call any tool through a standard interface, the argument goes, then the dashboard becomes obsolete. You stop opening tabs and clicking through reports. You ask your AI assistant a question, and it queries 12 different systems in parallel, pulls back the data, reasons across it, and gives you the answer.

This is the post-dashboard era. The interface moves from screens to conversations. Software stops being something you look at and starts being something that thinks for you.

The vision is correct. The timing is not.

The thing that gets glossed over

For MCP to deliver the future people are selling, one assumption has to hold: that the application on the other end of the protocol has a coherent, unified data model the model can reason against.

This assumption is almost never true.

Most software products in 2026 are not built around a coherent data model. They are built around a user interface that sits on top of a mess. The mess is usually a collection of database tables that were added incrementally over years, each one solving an immediate need, each one defining the same concept slightly differently.

In a typical SaaS product, ask three different parts of the system what a "customer" is and you will get three different answers. Ask what an "order" is, what "revenue" is, what "this month's performance" is, and the disagreement compounds.

Humans have learned to work around this. We squint at the dashboard, mentally apply the correction we know is needed, and trust our gut to fill the gaps. We are very good at this because we have to be.

A language model cannot do this. It will read whatever data is presented to it as if it were truth. It will reason confidently across inconsistencies it cannot detect. It will produce answers that sound right and are wrong.

I wrote about this in January in Context Graphs. The short version: AI without a structured data foundation is not analytics. It is hallucination at scale.

MCP does nothing to fix this. MCP just makes it faster.

The application problem

Consider what happens when you bolt MCP onto a product that does not have a unified data foundation.

The model asks: "What was our revenue last week?"

The application responds with whatever its "revenue" endpoint returns. But that endpoint is reading from three different tables, two of which use different definitions of revenue (gross or net, including or excluding refunds, currency-converted on which date), and the model has no way to know which definition is correct for the question being asked.

The model returns a confident answer. The executive nods. The number is wrong by 8 percent.

This is not a hypothetical. This is what is happening right now in companies that have rushed to slap MCP servers onto their existing products. The infrastructure works. The protocol works. The data does not.

The hard part of building for the AI era was never the protocol. It was always the data.

Why dashboards do not die yet

Here is the contrarian position: dashboards are going to...