MCP unveils new roadmap, focusing on five areas including long-running processing and AI agent authentication.

The Model Context Protocol (MCP), an open standard for connecting AI applications with external data and tools in a common way, released a new roadmap on August 22, 2026. The roadmap goes beyond simple tool calls and identifies key areas for the next 6 to 12 months, including processing long-running AI agents, agent authentication, and mechanisms for efficiently handling a large number of tools.

The New MCP Roadmap | Model Context Protocol Blog

https://blog.modelcontextprotocol.io/posts/mcp-roadmap/

MCP is a common standard for AI applications to access external systems such as files, databases, and search services. By using MCP, instead of preparing separate connection methods for each service, the MCP server exposes data and functions as 'tools,' which can then be called from corresponding AI applications using a common method. When Anthropic released MCP in November 2024, it was positioned as a standard for connecting fragmented data sources with AI.

Anthropic proposes and open-sources 'Model Context Protocol,' a universal protocol for bridging data between existing apps and AI systems - GIGAZINE

On the other hand, as the applications of AI agents expand, there are increasingly more situations where they need to run processes that take several minutes or more in the background, receive progress updates, or provide additional instructions to processes that are already running. A system that simply 'calls an external tool once and receives the results' is no longer sufficient. For enterprise use, mechanisms to distribute the load by using multiple servers and secure authentication methods for the AI agents themselves become necessary.

MCP underwent a major change with its specification revision on July 28, 2026. Previously, it relied on a mechanism to maintain sessions between the client and server, but the protocol eliminated the session and initialization exchanges, making the core of the protocol stateless and shifting to a method where each request can be handled independently. This means that even if multiple MCP servers are deployed behind a load balancer, it is no longer necessary to maintain a connection to a specific server, making it easier to operate MCP servers in a way similar to typical web services.

The common standard 'MCP' for connecting AI and external tools receives its biggest update ever: What has changed with the elimination of sessions? - GIGAZINE

The new roadmap is based on the July 2026 specifications and sets out five priority areas. The first, 'Basic Messaging Functionality for AI Agents,' outlines a policy to streamline previously separate mechanisms such as 'Tasks' for handling time-consuming jobs and 'subscriptions/listen' for receiving updates, making them usable in a consistent manner. They are also considering mechanisms to notify completion from the server side using webhooks and channels, rather than relying solely on polling where the client repeatedly queries. Tasks are currently an extension at the time the roadmap is released, but it is stated that the goal is to integrate them into the MCP core once they mature.

The second area of focus is the unification and strengthening of HTTP-based communication methods. While the remote MCP server has a configuration closer to a typical HTTP service as of July 2026, the locally running MCP server still uses 'stdio,' which utilizes standard input/output. To unify communication methods as much as possible, the MCP side is considering ways to handle Streamable HTTP on stdio as well, and has also proposed using HTTP/2 on standard input/output. In addition, there are plans to incorporate ETags, which are used to determine cache updates, to efficiently determine whether the retrieved results have been updated.

The third area of focus is AI agent identity management and security for enterprise use. MCP authorization has been based on the assumption that a human opens a browser and grants access, but with agents operating autonomously in the cloud, human intervention is not always possible, and there are also scenarios where only limited permissions are granted from a parent agent to a child agent. To reduce reliance on manually configured API keys and long-term valid refresh tokens, MCP is considering a mechanism that can verify the agent's own identity and delegated permissions by combining DPoP, which links access tokens to specific keys, Workload Identity Federation, which verifies the identity of workloads using an external identity infrastructure, and token exchange.

The fourth area of focus is improving the basic functions of MCP, such as tool invocation. One challenge identified is the burden on the MCP server as the number of tools it handles increases. In the system as of the time the roadmap was released, connecting to a server that publishes 100 tools would mean passing a large amount of tool information to the AI model even before the user asks a question, potentially increasing the amount of information the model has to handle and reducing the accuracy of tool selection. At the time the roadmap was released, they are considering 'Progressive Discovery,' which would present necessary tools and resources in stages, exploring a method that initially presents only small pieces of information and expands the options according to the content of the conversation. They also plan to standardize the format of tool execution results to reduce the problem of different clients handling the results.

The fifth focus area is improving the developer experience for Software Development Kits (SDKs). The plan is to generate SDK candidates and sample code for implementation from the specifications based on the specifications and conformance tests, and to verify whether this reduces the burden of manually keeping up with specification changes. The generated SDKs will be subjected to conformance tests to investigate which parts should be created using conventional code generation and which parts should utilize AI models.

The roadmap released this time outlines the priorities that MCP's core maintainers and working groups will focus on. Specification Enhancement Proposals that align with the roadmap will be given priority for review, but the content may change or be postponed depending on future discussions. MCP is expanding its scope to include processing and notifications for long-running AI agents, agent ID management, and large-scale server operations, and the MCP team says it will develop future specifications while soliciting participation from the community in each priority area.