97 million installs is not a developer experiment. That number, reported at the MCP Dev Summit in New York on April 2-3, 2026, describes infrastructure adoption. The announcement that followed confirmed what the install count was already signaling: Anthropic donated MCP to the Linux Foundation's newly formed Agentic AI Foundation (AAIF). MCP is no longer a protocol one company controls. It is a standard that an industry is building on.
The conventional read is celebratory. Open standards win, ecosystem grows, everyone ships faster. That read is accurate but shallow. The sharper implication is for the proprietary agent SDKs: Claude Agent SDK, OpenAI Agents SDK, and anything else that routes agent behavior through a vendor's abstraction layer. Those products just moved from "preferred path" to "optional add-on."
What the AAIF Actually Changes
The Linux Foundation does not host experimental projects. It hosts infrastructure. The Foundation's portfolio includes the Linux kernel, Kubernetes, and OpenTelemetry. Each became the default precisely because enterprise buyers needed vendor-neutral governance before committing procurement dollars. AAIF follows the same logic. MCP, contributed alongside participation from multiple labs, now has the governance structure that enterprise architecture committees require.
That governance shift matters more than the install count, though the install count matters too. 97 million installs means MCP compatibility is already a baseline expectation for any tool or model that wants to be taken seriously in an agent workflow. Before AAIF, a CTO could still argue that MCP was Anthropic's protocol, steerable by Anthropic's roadmap. After AAIF, that argument is gone. The protocol is owned by a foundation with multi-vendor participation. Betting against it now means betting against Linux Foundation governance, which enterprises have not historically been willing to do.
The Disintermediation Nobody Is Talking About
Here is the inversion the hype cycle misses. Most commentary frames MCP's growth as good news for the labs because it expands the tooling surface each lab's models can reach. That framing treats MCP as a feature multiplier for proprietary models. It is actually a substitution vector.
When an enterprise builds a multi-vendor agent fleet (the stated goal of most enterprise AI architecture work right now), they need an orchestration layer that does not privilege any single model provider. Claude Agent SDK routes through Anthropic's abstractions. OpenAI Agents SDK routes through OpenAI's. Standardize on either and you have picked a vendor at the orchestration layer, not just the model layer. That is the lock-in that procurement and security teams have been explicitly told to avoid.
MCP, governed by AAIF, is neutral at the orchestration layer. An enterprise that standardizes on MCP-compatible tooling can swap the underlying model provider without rebuilding their tool integrations. That is not a theoretical benefit. It is the architectural requirement coming out of every enterprise AI governance conversation I have seen in the past six months.
The proprietary SDKs will argue they provide capabilities MCP cannot: richer state management, tighter evals integration, better observability. Those arguments are real but temporary. Open standards absorb capabilities over time. OpenTelemetry absorbed most proprietary observability instrumentation approaches within three years of serious adoption. Kubernetes absorbed proprietary container orchestration faster. The pattern is consistent.
What Practitioners Should Do Now
If you are building agent tooling today and you are not exposing MCP-compatible endpoints, you are adding friction to every enterprise sale you will try to make in the next 18 months. That is a concrete business problem, not a philosophical preference for open standards.
If you are building internal agent infrastructure, the AAIF governance structure gives you the procurement justification to standardize on MCP now rather than waiting for further ecosystem maturity. 97 million installs plus Linux Foundation governance is a sufficient signal for most architecture decisions.
If you are evaluating a proprietary agent SDK, ask one question: what does the migration path look like when you want to add a second model provider? If the answer is "rebuild significant portions of your orchestration logic," you are looking at a lock-in tax, not a capability advantage.
The MCP Dev Summit in April 2026 will probably be remembered as the moment the agentic AI layer clarified. The protocol did not become important on April 2nd. It had already become important at 97 million installs. What April 2nd added was the institutional structure that removes the last credible argument for waiting.
Neutral infrastructure always wins the enterprise layer. The labs built the models. The Foundation now holds the protocol. Those are different businesses, and only one of them scales without a vendor relationship.