The integration treadmill

If you have wired a language model into a real product, you already know the tax. The model is brilliant in the abstract and useless in the specifics: it cannot see your database, your ticketing system, your file system, or last quarter's spreadsheet. Anthropic named the problem precisely when it introduced MCP — models are "constrained by their isolation from data, trapped behind information silos and legacy systems," and "every new data source requires its own custom implementation."

So we build connectors. One for Postgres, one for GitHub, one for Slack, each glued to a single model's tool-calling format. Swap the model and you rewrite them. Add a model and you do it all again. The community has a name for the shape of this mess: N models times M tools gives you N×M bespoke integrations. A shared protocol collapses that to N+M.

Why a protocol, not a framework

This is the distinction that matters, and it is why I keep telling people MCP is not “just another framework.” A framework is a way to write your code; you adopt it, and your code lives inside it. A protocol is an agreement about how separate programs talk to each other, no matter how each one is written. Frameworks compete and churn. Protocols, once enough people agree on them, quietly become the ground everyone else builds on — think HTTP, or SQL.

MCP's own documentation reaches for a hardware analogy: "Think of MCP like a USB-C port for AI applications." One standardized connector, and any compliant device plugs into any compliant port. You stop building adapters and start building features.

The closer precedent, though, is the Language Server Protocol. The spec says so outright: "MCP takes some inspiration from the Language Server Protocol, which standardizes how to add support for programming languages across a whole ecosystem of development tools." Before LSP, every editor needed a bespoke plugin for every language — an N×M mess. After LSP, a language author wrote one server and every compliant editor lit up. MCP is making the same bet for AI applications and the tools they reach.

The adoption signal

Standards live or die on who shows up, and MCP's first eighteen months are hard to argue with. Anthropic open-sourced it in November 2024. In March 2025, OpenAI adopted it — Sam Altman wrote that "people love MCP and we are excited to add support across our products." In April, Google DeepMind followed, with Demis Hassabis calling it "rapidly becoming an open standard for the AI agentic era." Microsoft went broad at Build in May 2025. And in December 2025 the protocol was donated to the Linux Foundation's new Agentic AI Foundation, alongside founding projects from Block and OpenAI.

When the three labs that build the frontier models all decide to speak the same integration language, that language stops being a bet and starts being infrastructure.

What this actually buys you

Concretely: you write a tool server once, and it works in Claude Desktop, in Claude Code, in an agent you build with the Agent SDK, and in clients from vendors who have never heard of you. What you build accrues value instead of decaying with the next model release.

That is the whole pitch. If you are connecting AI to the real world, learn the protocol first. The frameworks will come and go on top of it — the protocol is the part worth betting a career on.