MCP Explained: Why OpenAI, Google, and Microsoft Have All Adopted Anthropic's Invention in Six Months

The Problem Everyone Was Experiencing Without Naming It

Your AI assistant is brilliant but has no memory. It drafts a perfect email but can’t send it. It analyzes your content strategy but has never seen your editorial calendar. It advises you on your finances, but your bills are locked away in software it can’t access. Every tool in your business is an island, and the AI is left on the dock.

Until the end of 2024, the only way to connect an AI to a tool was to code a custom integration—for every model-tool pair. Ten models, twenty tools: two hundred potential integrations to write and maintain. Developers call this the N×M problem, and it’s exactly the situation with phone chargers before USB-C: a drawer full of incompatible cables. The comparison isn’t mine—the protocol’s official documentation describes itself as “a USB-C port for AI applications.” This is the problem that the Model Context Protocol, or MCP for short, was designed to solve. And the speed at which the industry has adopted it is, to my knowledge, unprecedented in recent web history.

Eighteen Months: From In-House Project to Global Standard

The timeline is worth recounting, because it explains why you can build on it with confidence. On November 25, 2024, Anthropic (the company behind Claude) released MCP as open source, with ready-to-use connectors for Google Drive, Slack, GitHub, and Postgres. Just another protocol, one might have said—the web’s graveyards are full of them.

Except that on March 26, 2025, the unthinkable happened: OpenAI adopted the standard of its direct rival. Sam Altman announced it in a tweet—“people love MCP.” Two weeks later, on April 9, Demis Hassabis confirmed that Google DeepMind was integrating Gemini. In May, Microsoft announced native support in Windows 11 at its Build conference. Three giants locked in a fierce commercial war had just agreed on a common standard. And on December 9, 2025, Anthropic transferred the protocol to the Agentic AI Foundation, under the umbrella of the Linux Foundation—co-founded by Anthropic, Block, and OpenAI, with AWS, Google, Microsoft, Bloomberg, and Cloudflare as platinum members. MCP no longer belongs to anyone, which means it belongs to everyone.

Figures from mid-2026 illustrate the scale: more than 20,000 MCP servers listed in the PulseMCP directory as of July 2026, 97 million monthly downloads of the Python and TypeScript SDKs as of March 2026, about ten official SDKs, and more than 300 client applications. For the record, the “ChatGPT plugins”—OpenAI’s proprietary attempt from 2023—were shut down in April 2024, after barely a year. The lesson: closed ecosystems die; neutral protocols survive. It was the Language Server Protocol that unified code editors; MCP is doing the same for AI.

In a nutshell: three terms to understand, not thirty

You don’t need to read the spec to grasp the architecture. Three roles. The host is the AI application you’re using—Claude, ChatGPT, Cursor. The MCP server is the adapter exposed by a service—Stripe, Figma, Shopify, your database. Between the two, the protocol defines a common language: the server announces “here’s what I can do” (search for a file, create an invoice, read data), and the host can use it without either party ever having heard of the other. An MCP server written once works with Claude, ChatGPT, Copilot, and all the others. N × M becomes N + M.

What really won me over was seeing it in everyday use. This blog you’re reading is powered by MCP: my posts live in a Supabase database, and when I work with Claude, it reads from and writes to it directly—no copy-pasting, no exporting. My Ahrefs data comes in via MCP. My Figma mockups, too. The line between “chatting with an AI” and “working with an AI” lies exactly there: real-world context, connected in real time.

What an SME Can Set Up Today, Without a Developer

The misconception to dispel: MCP is no longer just for programmers. Since July 14, 2025, Claude has offered a one-click connector library—Notion, Canva, Stripe, and about fifty others as of February 2026. ChatGPT followed suit with its developer mode in September 2025. On the services front, here are the launches that matter for a local business: Figma opened its MCP server in June 2025 (and as of February 2026, the connection is bidirectional—the AI can write to the canvas, not just read from it); Canva launched its own in July 2025, and this engine has been powering Claude Design since January 2026—I’ve documented this entire workflow in my article on importing Figma design systems. Shopify has enabled a default MCP endpoint on every store since the summer of 2025. Stripe exposes payments and invoices at mcp.stripe.com.

Translated into real-world scenarios: “Claude, show me the five best-selling products of the month and create three Canva visuals to promote them” is no longer science fiction—it’s just two checkboxes in the settings. For an SME, the question is no longer “Is this possible?” but “What do I connect first, and under what conditions?” And that’s where you need to slow down.

The Part That Enthusiastic Articles Don’t Mention

Connecting an AI to your actual tools is like handing it the keys. The year 2025 provided three documented warnings that I share with every client. In June, Asana shut down its MCP integration for two weeks: a bug was exposing one client’s data to other clients. In July, Simon Willison documented the textbook Supabase case: an attacker had slipped instructions into an ordinary support ticket, and the AI agent reading the tickets—connected to the database with administrative privileges—obeyed and exfiltrated the table of secret tokens. Willison gave this combination a name: the “lethal triad”—private data, untrusted content, and an output channel. When all three come together, prompt injection becomes a weapon. And in September, the first malicious MCP server was confirmed: postmark-mcp, an npm package that, in version 1.0.16, added a hidden copy of all your emails to the attacker’s server.

The overall figures aren’t any more reassuring: Equixly’s analysis found that 43% of the MCP servers tested were vulnerable to command injection, and Trend Micro identified 492 servers exposed on the Internet without any authentication. The spec has stepped up its game in response—mandatory OAuth for remote servers starting with the March 2025 release, user confirmations (“elicitation”) since June 2025, and ongoing hardening through the November 2025 release — but the protocol can’t protect you from a server you should never have installed in the first place.

My Verdict as a Practitioner

MCP is the safest infrastructure bet in the AI ecosystem in 2026: neutral governance under the Linux Foundation, unanimous adoption by major vendors, an ecosystem that doubles in size year over year, and an upcoming major release (finalized in late July 2026) already geared toward enterprise needs. For an SMB, my recommendation is a three-step process. Start with the official one-click connectors in the AI tool you’re already using—it’s free and reversible. Next, identify the integration that would save you the most time—for my clients, it’s almost always accounting, inventory, or the content calendar. And only then, if a connector is missing, consider a custom server: it takes just a few days of development, especially now that the standard is stable.

The AI of 2023 impressed with conversation. The AI of 2026 gets the job done, because it’s connected. MCP is the connector—and for once in the history of technology, everyone has agreed to use the same one.

Sources

• Announcement of the Model Context Protocol — Anthropic, November 25, 2024
• OpenAI Adopts Its Rival’s Standard — TechCrunch, March 26, 2025
• Google DeepMind Adopts MCP — TechCrunch, April 9, 2025
• Launch of the Agentic AI Foundation — Linux Foundation, December 9, 2025
• Figma MCP Server — Figma, June 4, 2025 · Canva × Claude — Canva Newsroom, January 2026
• The “Deadly Triad” and the Supabase Incident — Simon Willison, July 6, 2025
• Malicious postmark-mcp server — Snyk, September 2025
• MCP Security Statistics 2026 — Practical DevSecOps
• PulseMCP Directory (20,000+ servers, July 2026) · Official Documentation