MCP: when a protocol becomes infrastructure

In November 2024, Anthropic open-sourced the Model Context Protocol (MCP), a standard for connecting AI agents to external systems: databases, business software, workflows, tools.

In December 2025, MCP changed status. Anthropic handed it over to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation, co-founded with OpenAI and Block, and backed by Google, Microsoft, AWS, Cloudflare and Bloomberg.^[1][2]

It is no longer Anthropic's protocol. It is neutral industry infrastructure.^[3]

What MCP actually does#

MCP standardizes the way an AI agent connects to an external system.^[4]

Before MCP: every integration was a project in itself. A custom API, a security layer to reinvent, a data flow to tame, for every tool, for every system.

With MCP: you build once, connect everywhere. The protocol defines a common interface between the model and whatever it plugs into.

In numbers, at the time of the announcement:

• 10,000+ MCP servers published in one year
• 97 million monthly SDK downloads
• Natively supported by Claude, ChatGPT, Copilot, Cursor, and most of the major agentic tools

Why standardization changes everything#

A standard is about more than “simplifying connections”.

What HTTP made possible was not just loading pages, it was building entire systems on top of it: secure, auditable, scalable, interoperable.

What IoT did to connect physical objects to the internet, MCP is now doing to connect intelligence to systems, in the broad sense: business software, databases, infrastructure, human and machine workflows.

Standardization unlocks three concrete things:

1. Scoping: clear rules on how an agent interacts with a system, what it can do, what it cannot do.
2. Security: a defined attack surface, auditable and governable. Not a handcrafted integration every time.
3. Fluidity: build once, connect everywhere. Integration complexity stops being the number one obstacle to deploying AI in the enterprise.

The real shift#

The question will no longer be “is your AI powerful?”

But: “is it plugged in?”

An isolated model, however capable, remains a gadget on the periphery. What makes an AI agent valuable in the enterprise is its ability to cut across the organization, to read, act and coordinate across existing systems.

MCP lays the first brick of that world. Not the last: open questions remain around fine-grained security, access governance and agent identity management. But the foundation is there, neutral, open, and backed by the whole industry.

This is not a technical detail. It is a paradigm shift in what an AI integration is.

The question will no longer be “is your AI powerful?”, but “is it plugged in?”