The Model Context Protocol Just Quietly Became Core Infrastructure

Anthropic reported on Wednesday that the Model Context Protocol has crossed 97 million installs across major providers. That number came from a combination of its own telemetry, reports from Google and OpenAI about tool adoption in their developer products, and counts from independent server registries. The milestone is big enough to change the conversation about how AI tools get built.

For those who have not been paying attention, MCP is a protocol published by Anthropic in late 2024 that defines how AI models talk to external tools, data sources, and services. Think of it as the HTTP of AI agents. Before MCP, every company that wanted to connect a model to a tool had to invent their own protocol, often tied to a specific vendor's quirks. MCP made that step boring.

Boring protocols are how infrastructure actually gets built. HTTP is boring. SMTP is boring. SQL is boring. The protocols that lasted are the ones nobody had to think about anymore. MCP is on that track. The 97 million install number is less important than the fact that every serious AI provider now ships MCP-compatible tooling, which means a developer building an agent does not have to choose a vendor before choosing a protocol.

Why does this matter for operators? Because protocol standardization changes your architecture choices. Two years ago, if you wanted your internal data accessible to AI agents, you had to build a custom integration for each model vendor. Today, you build one MCP server, and every model that speaks MCP can access it. The integration cost just dropped by a factor of three or four, depending on how many vendors you care about.

The practical move for any team with internal systems worth exposing to AI is to stop building vendor-specific integrations. Build an MCP server. Let the agent on whichever model choose how to use it. Your integration work compounds across every vendor's model you adopt later, instead of needing to be redone each time you switch.

There is a second-order effect that matters more for the industry. With a shared protocol, the barrier to switching vendors drops. If your entire agent architecture talks MCP, switching from Claude to GPT to Gemini is closer to a deployment change than an architectural rewrite. Vendors know this. It explains why all of them adopted MCP quickly even though it is an Anthropic-originated standard. Not adopting would have meant losing the developer mindshare to competitors who did.

Why aren't we treating this as a bigger story? Because protocols are boring by design. The press loves model releases and benchmark wars. Nobody writes a splashy headline about a protocol adoption milestone. That is a mistake, because the protocol war is often where the real platform dynamics get settled, and MCP has clearly won.

For operators, the near-term roadmap implication is that your AI architecture should assume MCP as the default integration layer. Any internal tool you want exposed to an agent should live behind an MCP server. Every public service worth integrating is probably already available as an MCP server, and if it is not, building one is a weekend project.

The longer-term implication is that the center of gravity in enterprise AI is shifting from 'which model' to 'which tools and which data.' The model is increasingly a commodity. The unique value of your AI stack is in the tools you expose, the data you connect, and the workflows you assemble. MCP makes that value portable across vendors, which means it accumulates to the company building the workflow rather than to whichever model is currently fashionable.

The quiet arrival of MCP as a default is exactly the kind of thing that separates infrastructure winners from infrastructure losers in hindsight. Anthropic did not try to make MCP proprietary. They made it easy to adopt. They let competitors embrace it. They captured developer mindshare through openness rather than exclusivity. That playbook worked for Google with Kubernetes, for Red Hat with Linux, and for many others before. It is now working for Anthropic with MCP.