The next iteration of the Model Context Protocol (MCP) will enable IT teams to deploy stateless servers that will make it easier to deploy artificial intelligence (AI) applications at higher levels of scale.
Speaking at an MCP Dev Summit North America conference in New York today, David Soria Parra, a member of the technical staff at Anthropic who is also co-creator of MCP, told conference attendees the next iteration of the specification, due out in June, has been developed in collaboration with Google and Microsoft to enable cloud service providers to more easily spin up MCP servers on demand.
At the same time, the technical oversight committee for MCP, now being advanced under the auspices of the Agentic AI Foundation (AAIF), is also working on a task capability that will make it easier to run long-running autonomous workflows versus continuing to rely on a request and response mechanism that would otherwise need to be reinvoked multiple times.
Additionally, maintainers of the MCP project are working on adding a triggers capability that will enable servers to initiate an action versus always having to rely on an MCP client, added Soria Parra.
Support for retry semantics, expiration policies, native streaming and reusable skills that are based on domain knowledge are additional capabilities that are expected to be added in 2026.
Finally, updates to the Python and TypeScript software development kits (SDKs) will provide access to more efficient MCP clients and servers, said Soria Parra.
In general, IT teams should be taking advantage of composable techniques such as progressive discovery that only loads tools at runtime to create more efficient clients, noted Soria Parra.
Overall, MCP SDKs are now being downloaded 110 million times a month, a number that will only continue to exponentially increase as enterprise IT teams connect more MCP servers to systems of record applications residing behind firewalls in enterprise IT environments, added Soria Parra. “The reality is this is the year we will see agentic systems make a big impact,” he said.
Longer term, the maintainers of the MCP project are also exploring how best to secure AI agent and MCP server interactions using, for example, extensions to the protocol, noted Soria Parra.
The challenge is that as the number of use cases for MCP continues to expand, the need for MCP expertise continues to rise. Server-initiated triggers, streaming tool results, cross-app identity federation, and reusable skills are all on the near-term horizon, said Mitch Ashley, vice president and practice lead for software lifecycle engineering at the Futurum Group.
MCP is establishing itself as the connective layer for agentic systems, he added. “Teams building on MCP today need to architect around those gaps now,” said Ashley.
Each IT team depending on their internal expertise and appreciation for cybersecurity concerns will need to determine at what pace they will deploy MCP servers. The one thing that is certain is that MCP servers will soon be strewn across the enterprise as organizations look to provide AI agents with access to the data and context needed to generate the most accurate outputs possible.
