When Anthropic quietly released the Model Context Protocol in November 2024, most developers treated it as an interesting Anthropic-specific experiment. Seventeen months later, it has crossed 97 million installs, earned governance under the Linux Foundation, and been adopted by every significant player in the AI industry. The protocol is no longer an experiment — it is infrastructure.
What MCP Actually Does
Model Context Protocol defines a standard interface for how AI agents communicate with external tools, data sources, and APIs. Before MCP, every agent-to-tool integration required custom code: a bespoke connector for a database here, a hand-rolled API wrapper for a CRM there. The result was a fragmented ecosystem where tooling written for one agent rarely worked with another.
MCP changes this with a simple premise: any MCP-compatible agent can use any MCP server without custom integration code. The analogy that has stuck in the developer community is accurate — it is effectively USB for AI tools. Plug in a server, and any compliant agent can immediately read from it, write to it, and call its functions.
The ecosystem that has grown around this standard is substantial. Over 5,800 community and enterprise MCP servers are now publicly available, covering databases, cloud platforms, CRM systems, developer tools, analytics services, and a long tail of specialized integrations. A developer building an agentic workflow today rarely needs to write a connector from scratch.
From Anthropic Lab to Linux Foundation
The protocol’s governance trajectory is what distinguishes it from other would-be AI standards. In December 2025, Anthropic donated MCP to the newly formed Agentic AI Foundation under the Linux Foundation umbrella — the same organization that stewards Linux, Kubernetes, and other foundational open-source infrastructure.
The founding membership list reads like a roll call of the industry’s largest stakeholders: OpenAI and Block as co-founders, with AWS, Google, Microsoft, Cloudflare, and Bloomberg as platinum members. The donation was strategically important. By ceding ownership to a neutral foundation, Anthropic removed the most credible objection competitors had to adopting the standard — namely, that standardizing on MCP meant standardizing on Anthropic’s terms.
The result has been near-universal adoption. OpenAI, Google DeepMind, Microsoft, and Meta all now ship MCP-compatible tooling. What began as one company’s internal protocol is now the mechanism by which agents across the entire industry connect to external systems.
Why the 97 Million Number Matters
Raw install counts can mislead — downloads are not the same as deployments. But the 97 million figure, crossed in March 2026, is notable less for its size than for what it represents structurally.
Previous AI infrastructure standards wars — around model formats, inference APIs, embedding interfaces — tended to fragment. Multiple competing standards persisted in parallel, forcing developers to pick sides or maintain compatibility shims. MCP’s adoption curve looks different: one standard, converging adoption, no credible rival.
This is partly a timing effect. The agentic AI wave hit at precisely the moment MCP had gained enough critical mass that building against a competing standard carried real opportunity cost. Developers who wanted their tools to work across Claude, GPT-4, Gemini, and open-source models had one obvious path — and it ran through MCP.
The Infrastructure Layer Is Locked In
For enterprise buyers evaluating agentic AI deployments, the MCP milestone carries a practical implication: the integration layer has stabilized. Tools built on MCP today are not bets on a provisional standard — they are bets on infrastructure that has already won.
For developers, it means the commoditized problem of agent-to-tool connectivity is largely solved. The competitive surface has shifted upstream, to the quality of the agents themselves and the intelligence of the workflows they execute.
Anthropic released a protocol. The industry turned it into a foundation. That transition — from vendor experiment to neutral standard — is one of the quieter but more consequential shifts in the AI stack so far this year.
