MCP Dev Summit 2026: the protocol grows up for enterprise use

For a few months, social media has been full of posts declaring that the Model Context Protocol is already past its peak. The MCP Dev Summit in New York on 2–3 April told a different story. Around 1,200 people came to the first summit since MCP was handed over to the Agentic AI Foundation at the Linux Foundation in December 2025, and much of the programme was about running MCP in production, not trying it out.

For integration teams, the interesting part is not the hype cycle. It is that the protocol is now being shaped around the problems enterprises actually have: scale, long-running work and identity.

What happened in New York

The Agentic AI Foundation, which now hosts MCP, said it has added well over a hundred members since it was formed. Several large companies showed MCP in production. AWS responded directly to the “MCP is dead” posts by saying that is not what it sees inside Amazon. Uber described running tens of thousands of agent executions a week through its own MCP gateway and registry, and Duolingo, Datadog and Bloomberg also presented production use.

The core maintainers outlined three priorities for the next specification release, planned for June:

• Stateless transport, so MCP servers can run on serverless platforms without keeping a session open for each client.
• Long-running tasks, for jobs that take minutes or hours instead of seconds. An experimental version arrived in the November 2025 specification.
• Enterprise authentication through the organisation’s own identity provider, so users do not have to approve access app by app.

A new working group is also looking at triggers and events, so MCP servers can notify agents when something happens instead of waiting to be asked.

Why it matters for integration

All three priorities are familiar to anyone who has run an integration platform. Stateless design is what makes it possible to scale out. Long-running tasks are the difference between a demo and a real order or onboarding process. And central identity is what security teams ask for first.

Uber’s setup is also worth noting. A large MCP deployment looks a lot like API management: a gateway in front, a registry of what is available and central rules. That is the same pattern the iPaaS vendors are building into their platforms, and it suggests MCP servers will end up being managed much like APIs are today.

How it compares

MCP is not the only open protocol for agents. A2A, which connects agents to other agents, is also a Linux Foundation project. The two are increasingly described as complementary: MCP connects an agent to tools and data, while A2A connects agents across teams and platforms.

For buyers, the practical question is less about which protocol wins and more about how each platform supports both, and how quickly vendors adopt new specification versions.

What to watch

1. The June specification. Check when your integration platform supports stateless transport and tasks. Older MCP servers may need changes.
2. Identity. Ask vendors how their MCP servers will work with your identity provider, and whether users can act with their own rights.
3. Registry and gateway. If you plan more than a handful of MCP servers, decide early where they will be registered and how traffic will be controlled.

Read more in our guide MCP explained.