The Model Context Protocol's core maintainers, David Soria Parra and Den Delimarsky, published a roadmap outlining five priorities for the protocol's next specification release, according to a post on the MCP blog.
The first priority adds agentic messaging primitives, including server-initiated events and a matured Tasks extension, for work that does not fit a single request-and-response exchange. "Modern agentic workloads no longer fit the standard request-and-response pattern," the maintainers wrote, pointing to a need for streaming results and mid-flight work steering.
The second priority moves every deployment mode onto HTTP-based transport, including local servers, which currently run over stdio and would shift to Streamable HTTP instead.
A third priority replaces manually managed API keys with standardized agent identity, using Demonstrating Proof of Possession (DPoP) and Workload Identity Federation, so a server can verify which agent is calling it rather than just which key it holds.
The fourth priority standardizes tool result handling and adds progressive tool discovery, meant to cut model overhead when an agent has access to a large toolset. The fifth invests in software development kit ergonomics, specification conformance testing and documentation across languages, the maintainers wrote.
MCP has become the default way AI agents connect to external tools and data since Anthropic introduced the protocol in late 2024, so these changes ripple through every server and client built on it. Builders running an MCP server should watch the identity and transport items closest: replacing API keys with identity federation is the kind of change that breaks integrations quietly if it ships without a migration path.