Model Context Protocol: How MCP Became AI's Universal Connector

A common plug for AI

As AI assistants take on more practical work, one recurring problem has been how to connect them to the systems where real information and real work live. The Model Context Protocol, or MCP, is the industry's most visible answer. It is an open-source standard that defines how AI applications link to external systems 1. Through MCP, assistants such as Claude or ChatGPT can reach data sources like local files and databases, tools like search engines and calculators, and workflows such as specialized prompts. That access lets them retrieve key information and carry out tasks rather than just generate text 1.

The protocol's own documentation uses a hardware analogy. MCP is meant to be a USB-C port for AI applications. USB-C gave electronics a single standardized connector, and MCP aims to give AI software one standardized way to plug into outside systems 1. The comparison is marketing shorthand, but it captures the core value. Without a shared protocol, every pairing of a model and a tool needs its own custom integration. With one, a tool exposed once through an MCP server can in principle work with any compliant client.

From one company's project to neutral ground

The biggest governance change came in December 2025. Anthropic donated MCP to the Agentic AI Foundation (AAIF), a fund operating under the Linux Foundation 2. That move matters more than it might seem. Standards controlled by a single commercial vendor often struggle to win full commitment from that vendor's competitors. Placing MCP under a neutral foundation is a common route for turning a promising specification into durable shared infrastructure.

The timing also matters. The donation came after MCP had already gained traction with rivals, so neutral stewardship looks less like a bid for relevance and more like the formalization of a standard the industry had already chosen.

Adoption across competitors

The clearest sign of that came from OpenAI. In March 2025, OpenAI formally adopted MCP after building the standard into its products, including the ChatGPT desktop app 2. In September 2025 it extended support to ChatGPT apps, opening the door to third-party access inside ChatGPT 2. When one of the largest AI developers adopts a protocol tied to a direct competitor, it suggests MCP is solving a problem the industry would rather share than fight over.

Support extends well beyond model makers. MCP can be integrated with Microsoft's Semantic Kernel and with Azure OpenAI, and MCP servers can be deployed on Cloudflare 2. That breadth across developer frameworks, cloud platforms, and edge infrastructure is what turns a specification into an ecosystem.

Signs of scale

The numbers reported through 2026 point to production use, not experimentation. In April 2026, the AAIF held the MCP Dev Summit North America in New York City, which drew roughly 1,200 attendees 2. That same month, Salesforce's Headless 360 platform began routing customer and agent interactions through MCP. By late May, Salesforce said it had processed 4.5 million MCP calls since launch 2.

By mid-2026, more than 10,000 MCP servers had reportedly been deployed in production, and the protocol's SDKs were being downloaded over 97 million times a month 2. These figures are self-reported or aggregated rather than independently audited, so they deserve some caution. Even so, they describe a protocol that has moved into enterprise workflows.

Reading the picture

The official documentation and the adoption record tell complementary stories. The documentation offers the pitch: a universal connector that lets AI reach files, tools, and workflows 1. The adoption record shows that pitch being accepted by competitors, cloud providers, and enterprise software vendors, and then handed to a neutral foundation 2. The two accounts do not conflict. One explains what MCP is for, and the other shows that the industry has largely agreed to use it.

MCP looks to be on its way to becoming default plumbing for agentic AI, much as HTTP became for the web. The forces behind that are familiar: a simple, clearly explained purpose, early buy-in from rival platforms, and governance outside any single company's control. The open questions are the ones that arrive with success. These include how the foundation manages the specification's evolution, how security is handled as thousands of servers expose sensitive systems to AI agents, and whether the standard stays lightweight as more stakeholders press for features. For now, anyone building AI tools that need to touch real data or take real actions should treat MCP as the baseline.