Before MCP existed, connecting an AI model to an external tool was a custom engineering job. Every time. Connecting Claude to your company’s GitHub repository required one integration. Connecting it to Slack required another. A database, another. A CRM, another. Every new AI model your team evaluated, or every new tool your stack adopted, meant the same custom work again from scratch.
Anthropic called this the M×N problem: M models multiplied by N tools equals M×N separate integrations to build, maintain, and debug. As AI use in enterprises scaled through 2024, the integration tax was quietly becoming the thing that made serious AI adoption unscalable — not the models themselves, but the engineering effort to wire them to anything useful.
MCP was the answer to that specific problem. And it worked fast enough that even Anthropic’s direct competitors adopted it within months.
What Is MCP?
MCP — Model Context Protocol — is an open standard that gives AI models a single, standardized way to connect to external tools, data sources, and services which was developed by Anthropic and publicly released on November 25, 2024. Instead of a separate custom integration for every model-to-tool pair, each AI model implements MCP once and each tool or service implements MCP once therefore collapsing the M×N integration problem into M+N.
The simplest analogy is USB-C: before it, every device needed its own cable. After it, one connector works across everything. MCP does the same thing for AI integrations, not because the analogy is perfect, but because the underlying logic is identical. One standard, adopted on both sides, eliminating the need for proprietary adapters in between.
FURTHER READING
➤ AI in Business Processes: Use Cases, Benefits & Implementation
What Is an MCP Server?
The architecture behind MCP involves three roles, each with a specific job.
The MCP host is the AI application the user actually interacts with doubling as the environment where the AI model lives. Claude Desktop and Claude Code are Anthropic’s reference MCP host implementations, meaning they are designed from the ground up to connect to MCP servers. Cursor, VS Code with Copilot, and ChatGPT’s desktop app are other examples of hosts that have adopted the standard.
The MCP client is a component that lives inside the host. It manages the connection to a specific server — one client per server connection, kept isolated to simplify error handling and security.
The MCP server is the external programme or service that exposes capabilities to the AI. As Zuplo’s technical guide describes it, an MCP server receives structured requests from clients, executes the requested operations, and returns results. Each server can expose three types of capabilities: tools (functions the AI can call, like running a search or creating a ticket), resources (data the AI can read, like files or database records), and prompts (reusable templates for specific interaction patterns).
GitHub’s MCP Server, which entered public preview on April 4, 2025, is one of the most widely used examples: it exposes GitHub’s APIs — repositories, pull requests, issues, actions — as MCP tools, so any MCP-compatible AI model can read and interact with a GitHub codebase without a custom integration. Slack, Google Drive, Postgres, Jira, Linear, Snowflake, Stripe, and hundreds of other services have published their own MCP servers following the same pattern.
Anthropic’s official MCP documentation provides the full technical specification, server-building guides in TypeScript, Python, Java, Kotlin, Go, and several other languages, and a growing registry of available servers.
From Anthropic’s Idea to Industry Infrastructure
In March 2025, OpenAI who stands as Anthropic’s most direct competitor, officially adopted MCP across its Agents SDK, Responses API, and ChatGPT desktop app. CEO Sam Altman posted simply: “People love MCP and we are excited to add support across our products.” As technology analyst Shelly Palmer noted at the time: “Why would OpenAI adopt its rival’s protocol? Because standardization will speed AI adoption.” It was the clearest possible signal that MCP was solving a real infrastructure problem, not just serving Anthropic’s interests.
In April 2025, Google DeepMind CEO Demis Hassabis confirmed MCP support for upcoming Gemini models. In May 2025, at Microsoft Build 2025, GitHub and Microsoft joined MCP’s steering committee, with Microsoft announcing a preview of Windows 11 MCP integration. By December 2025, Anthropic donated the MCP specification to the Agentic AI Foundation — a new directed fund under the Linux Foundation, with Google, Microsoft, AWS, and Cloudflare as founding co-sponsors. MCP had stopped being “Anthropic’s protocol” and become neutral infrastructure.
The adoption numbers reflect that shift. By December 2025, MCP had more than 10,000 active public servers and 97 million monthly SDK downloads across its Python and TypeScript implementations — up from roughly 100,000 downloads at launch thirteen months earlier. By May 2026, the GitHub ecosystem showed over 15,900 repositories tagged with the mcp-server topic, with the core modelcontextprotocol/servers repository accumulating over 86,000 stars.
What MCP Means in Practice
The most concrete production result available comes from “Block” the parent company of Square and Cash App, which co-developed the MCP standard with Anthropic and built an open-source MCP-compatible AI agent called Goose for internal use. Block’s deployed Goose connects by default to Snowflake, GitHub, Jira, Slack, Google Drive, and internal APIs — all through MCP — turning the AI agent into a single interface for the entire tool stack. Reported employee outcomes: 50–75% time savings on common tasks, with work that previously took days completing in hours.
That result illustrates the structural shift MCP enables. The value of an AI agent in an enterprise context is not primarily the model’s intelligence — it’s the model’s access to the systems, data, and tools that make intelligence actionable. An AI agent that can only work with information inside its context window is useful. One that can read your database, query your GitHub repos, pull a Jira ticket, and draft a Slack message, all from a single prompt making this instance a different category of tool.
Before MCP, building that second type of agent meant custom integration work for every connection. With MCP, each connection is one standard implementation, reusable across any MCP-compatible model. The integration sprint that used to be a prerequisite for every new AI deployment is, increasingly, a solved problem.
For enterprises evaluating AI agents and agentic AI workflows, MCP is now effectively table stakes — the infrastructure layer that needs to be present before serious deployment begins, in the same way TCP/IP needed to be present before the internet could become useful.
The Take Away
MCP arrived quietly in November 2024 as an answer to a specific, tedious, expensive engineering problem: the cost of wiring AI models to the tools they need to be useful. Within thirteen months, it had been adopted by every major AI platform, donated to neutral governance under the Linux Foundation, and generated 97 million monthly SDK downloads.
The speed of that adoption is the signal. Standards that solve real infrastructure problems spread without needing marketing. MCP spread because the M×N integration problem was genuinely painful, and a clean M+N solution was genuinely valuable. The enterprise AI deployments that move fastest in 2026 are the ones that treated MCP as infrastructure from the start — not as a feature to add later.
