A behind-the-scenes protocol is quickly becoming one of the most important building blocks of the AI economy. Model Context Protocol (MCP) is gaining traction as companies race to deploy AI agents that can actually do things: book services, update systems, and run workflows across the internet.
With backing from major AI players and a recent handoff to the Linux Foundation, MCP is moving out of the experimental phase and into enterprise infrastructure, raising both excitement and serious questions about control, security, and governance.
For some, MCP will help enable the next generation of workflow systems (the "agent-based" systems) to work across applications without requiring each system builder to recreate an integration for each tool or model to achieve their goals.
However, others caution that there will be no substitute for strong controls over the implementation, deployment, and operation of MCP to ensure appropriate levels of security, governance, and reliability.
MCP is important because it was among the first infrastructure components developed to transition the Internet from primarily serving humans to primarily serving machines. This is notoriously tricky to achieve because human interaction on the Internet occurs through instinct and imagination, while machine interaction occurs through established rules and logic.
From APIs to agents, and why this feels familiar
For anyone who lived through the rise of APIs, MCP might feel familiar. But they are not the same thing. MCP is a protocol and tooling layer designed specifically so LLMs can discover and use tools (often backed by APIs), whereas APIs are the generic way any software talks to another system
APIs enable applications to communicate with one another programmatically, enabling mashups, platforms, and entire ecosystems of third-party development. MCP builds on that idea, but shifts the assumption about who is doing the talking.
MCP assumes an AI system deciding, in real time, which tool to use and how to use it. Instead of being designed once and deployed, integrations are discovered and selected dynamically by models that respond to intent. But not everyone is convinced.
Why read-only MCP breaks down in practice
Just as MCP moves into the corporate tech stack, so do the same old barriers to progress. For security teams, the primary reason to limit access is the ongoing emergence of new vulnerabilities. In response, many teams want to make all MCP connections read-only (only able to receive data; unable to take action).
Many see this as an incorrect assumption about how agents add value. Even if no harm can occur, no good can happen either. A read-only MCP connection can provide a summary of a CRM record, but it cannot update it, close a ticket, trigger a workflow, or complete an entire task. The result is usually extreme frustration.
When MCP implementations are overly restrictive, teams typically circumvent them. Teams will connect directly to APIs, establish their own MCP servers outside the governance framework, and implement their own AI. The goal of minimizing the risk of using MCP, in fact, increases it by creating fragmented access and visibility.
For IT leadership, this is the actual barrier to entry. MCP should be viewed as an infrastructure component that enables work while establishing appropriate guardrails based on the specific intent, context, and privileges of users accessing the system, rather than restricting it with broad rules.
As MCP enters the enterprise tech stack, the same old friction points return. Security teams want to restrict access for all the right reasons, given the ongoing emergence of vulnerabilities. The instinct is to make MCP connections read-only, allowing models to retrieve data but preventing them from taking action.
MCP as Internet plumbing, not an AI feature
By enabling new ways to interact with the Internet through capabilities rather than buttons or pages, MCP can fundamentally change how the plumbing works. Rather than navigating a website to find an action or workflow, Agents call on a tool or capability to perform the action.
Models can run a workflow without navigating a website's interface. The Web is now a surface for machines first, and humans provide the intent, rather than walking them through the steps.
We have been here before. Mobile apps didn't eliminate the web; they just made it a backend. Today, you may still enter a URL to get to where you're going, but the way you find out about where you're going has shifted from entering a URL directly to looking at App Stores & Feeds to discover where you can go next.
MCP is setting up a similar shift, but instead of being driven by platforms (i.e., Apple Store, Google Play), it will be driven by AI. But the moment MCP stopped looking like an experiment and started looking like infrastructure came when Anthropic donated it to the Linux Foundation.
Enterprise governance, beyond identity
Online services have traditionally been designed for human users. But service development is increasingly shifting toward AI agents using services for shopping, booking, customer support, and internal workflows (essentially bypassing the interface).
With each transaction executed independently, options compared, and outcomes returned by the AI agent, we are seeing a significant shift in how online discovery occurs. Discovery will no longer happen through search results; instead, it will shift toward tool catalogs and agent preferences, with important implications for traditional business models.
Many users are turning away from Google searches and the frustration of scrolling through a page of irrelevant sponsored results. As the shift towards AI continues, visibility will ultimately rely on the quality of the agent's integration into the system.
Reliability, execution success, trust signals, and AGO are more likely to determine ranking, rather than customer click-through rates and traditional SEO.
Despite the hype, we have been here many times before. App stores replaced browsers as gatekeepers for discovery. Social platforms reshaped distribution, and now MCP is raising similar questions about who controls access when agents mediate interactions.
If a few large-scale AI platforms come to dominate agent usage, they will also be able to dictate which services are available, prioritized, and ignored. This raises many of the long-standing issues regarding competition and market power. Sure, MCP is enabling this shift, but not causing it.
Why Critics Say MCP Is Overhyped
Not everyone is buying into the hype that surrounds all things MCP. Spend time in any tech forum or subreddit, and you won't be short of engineers and platform developers who say they don't think MCP really brings anything new to the table. To them, saying that MCP represents a breakthrough capability feels more like marketing inflation applied on top of the same old engineering work that has been going on for years.
Developers express concerns about "context-bloat," which occurs when dozens of MCP servers expose hundreds of tools that each must be loaded into the model's working context, consume tokens, and degrade performance before performing any actual work.
Why is MCP bigger than agent hype?
One of MCP's biggest strengths is its model-agnostic design. As more agents become increasingly capable, more services feel pressure to integrate with them. This feedback loop is how infrastructure changes become irreversible.
The average consumer will never hear the phrase Model Context Protocol or have any interest in it, and that's a good thing. Sure, they will notice that AI systems can do more, with fewer steps, across more services. When infrastructure disappears from view, it has usually won.
MCP is still early. But when a protocol gains cross-vendor support, neutral governance, and alignment with economic incentives, enterprises will inevitably hop on the hype train once again rather than risk getting left behind.
In 2025, Agentic AI and AI agents entered the mainstream conversation. 2026 looks set to be the year when MCP gets everyone talking about how its so-called plumbing is reshaping the Internet.
Unlock more exclusive Cybernews content on YouTube.
Your email address will not be published. Required fields are markedmarked