Model Context Protocol Outlines New Roadmap for AI Agents

The core maintainers and community working groups of the Model Context Protocol (MCP) have published an updated roadmap to guide the open-source standard over the coming months. A primary focus of the new plan is adapting to complex agent behaviors that go beyond traditional query-and-response interactions. To support long-running loops and mid-flight steering, the project is prioritizing server-initiated events like webhooks and channels alongside the maturation of the Tasks extension, known as SEP-2663, to integrate it directly into the core specification.

Following the 2026-07-28 release, which treated remote MCP servers like standard HTTP workloads, the roadmap aims to unify transport modes. Developers will soon be able to use Streamable HTTP over stdio for local servers, simplifying client and server creation. For enterprise security, the project is moving away from manual browser approvals toward automated agent identities. This security push includes finalizing Demonstrating Proof of Possession (DPoP) and implementing Workload Identity Federation, token exchanges, and the ID-JAG grant for Enterprise-Managed Authorization, while collaborating with the IETF OAuth and WIMSE working groups.

The roadmap also addresses developer pain points in tool calling and scaling. Currently, a tools/call response can return data in multiple formats, leaving server developers uncertain about what the client will display to the model. MCP plans to standardize this contract and introduce progressive discovery, allowing servers to reveal tools gradually rather than overwhelming models with massive catalogs. Finally, the maintainers are upgrading SDK ergonomics to ensure libraries are highly accurate for developers who use AI agents to write their code. Contributors can also test experimental features under the SEP-2133 framework.