From April 2 to 3, the MCP Developer Summit North America was held in New York, drawing 1,200 attendees and featuring over 95 sessions. This marked the first truly formal gathering since the Model Context Protocol established its own "annual conference."
The density of topics and the number of attendees make one thing clear: MCP is no longer just a developer toy.
From Protocol to Infrastructure
MCP was originally proposed by Anthropic in 2024 to solve the chaos of AI tool integration—every AI application and every tool had to be individually connected, at a very high cost. MCP provides a unified communication standard.
As of April 2026, the protocol's Python and TypeScript SDKs together exceed 97 million monthly downloads. Anthropic, OpenAI, Google, Microsoft, and Amazon have all adopted it. The Linux Foundation established a dedicated Agentic AI Foundation (AAIF) last December to maintain it.
The most closely watched technical proposal at the summit was SEP-1442, put forward by MCP co-founder David Soria Parra (Anthropic): changing the protocol from a stateful session architecture to a stateless request architecture.
This may sound like a low-level detail, but its impact is significant. A stateful protocol requires maintaining long-lived connections, which in enterprise environments leads to high operational costs and complex failure recovery. Switching to stateless means each request is handled independently, making horizontal scaling easier and simplifying integration with existing microservice architectures.
Real-World Deployments at Uber and Amazon
The most impressive part of the summit was not the protocol itself, but the real-world deployment data from several large companies.
Uber built an MCP Gateway and Registry, exposing its internal Thrift, Protobuf, and HTTP interfaces uniformly to AI agents. The entire platform is written in Go, and the Gateway layer performs PII redaction before requests are sent to external models. The number of agent executions running through this platform each week has already reached tens of thousands.
Amazon built an internal MCP tool discovery infrastructure and began packaging MCP tools, agent skills, and standard operating procedures (SOPs) into reusable configuration modules that different teams can directly adopt.
These two cases illustrate that MCP's adoption in large companies is not about one agent connecting to one tool. Instead, it involves a Gateway that centrally manages all agent tool calls, with a Registry serving as the control plane for the tool catalog. Docker, Kong, and Solo.io all follow the same approach, converging on an identical architectural pattern.
Claude Code Saves 85% of Tokens
One data point Anthropic presented at the summit deserves special mention: through progressive tool discovery, Claude Code reduced the token consumption caused by tool descriptions by approximately 85%.
The approach: if a tool's description would exceed 10% of the current context window, Claude Code defers loading that description until it is actually needed.
This is not just an engineering optimization; it reflects a design trade-off. When context resources are limited, it is better to lazily load tools than to let a large number of potentially unused tool descriptions fill the window. If the results are clear, this approach will likely be adopted by more MCP client implementations.
Ecosystem Numbers
As of the summit, AAIF membership exceeded 100 organizations, spanning multiple tiers. During the summit, the Linux Foundation also announced the formation of the x402 Foundation, dedicated to managing the payment layer protocol for AI agents.
MCP itself is positioned as the "access layer for tools and resources," the A2A protocol handles the "coordination layer between agents," and x402 manages the "payment layer." Together, these three layers are essentially building a complete foundational protocol stack for agentic AI.
That large companies are willing to run this stack in production environments is the most direct proof that MCP has evolved from a "concept protocol" into an "engineering standard."
Sources: CocoLoop; AAIF's MCP Dev Summit: Gateways, gRPC, and Observability Signal Protocol Hardening (InfoQ); Linux Foundation Announces the Formation of the Agentic AI Foundation (Linux Foundation)