MCPがセッションIDを捨ててステートレスになる2026-07-28仕様The MCP specification dated 2026-07-28 removes session IDs, shifting the…

The Model Context Protocol (MCP) is reportedly headed for a notable architectural change: according to a specification revision dated 2026-07-28, the protocol removes session identifiers in favor of a stateless design. Because MCP has become a widely used standard for connecting AI assistants to external tools and data sources, any change at the transport layer touches a broad ecosystem of server implementations. Existing developers will likely need to review and update their code to remain compliant.

MCP, originally introduced by Anthropic in late 2024, standardizes how applications expose tools, resources, and prompts to large language models using JSON-RPC 2.0 messages. The protocol defines multiple transports, most prominently a local stdio transport for process-to-process communication and an HTTP-based transport for networked deployments. In earlier revisions, the HTTP transport supported session continuity through an Mcp-Session-Id header: during initialization, a server could assign a session ID that the client would then include on subsequent requests. That mechanism let a server associate a sequence of calls with a specific client and preserve state between them.

Under the new specification, that session mechanism appears to be eliminated. Each request is intended to be self-contained, so any server instance should be able to handle any incoming request without relying on previously stored session context. The stated motivations are simpler server implementations and easier horizontal scaling. In practice, a stateless model removes the need to manage session lifecycles, storage, and expiration, which have historically been sources of complexity and subtle bugs. It also aligns MCP more closely with conventional REST-style and serverless deployment patterns.

The scaling argument is straightforward. When a server holds per-session state in memory, requests from a given client typically must be routed back to the same instance, a pattern often called session affinity or "sticky sessions." That requirement complicates load balancing and can leave traffic unevenly distributed. A stateless protocol lets operators place identical server instances behind a standard load balancer and add or remove capacity freely, since no single instance owns a client's context. This is particularly attractive for deployments on autoscaling platforms, container orchestration systems, and function-as-a-service environments, where instances are frequently created and destroyed.

The change is not without trade-offs, and developers should evaluate their specific use cases. Features that depended on server-maintained session state, such as long-lived subscriptions or server-initiated notifications tied to a particular connection, may require rethinking. Where continuity is still needed, that context likely has to be carried explicitly in each request or persisted in an external store such as a database or cache, rather than held implicitly on the server. In effect, statelessness does not remove the need for state so much as it shifts responsibility for managing it, moving it out of the transport layer and into either the client or a shared backing service.

For teams maintaining existing MCP servers, the practical impact depends on how heavily they relied on the session header. Implementations that already treated each call independently may need only minor adjustments, while those built around stateful sessions will require more substantial refactoring. It is worth confirming details against the official specification and reference software development kits rather than relying on summaries, since transport behavior, backward-compatibility provisions, and any transition guidance are defined there.

This revision fits a broader industry trend. MCP has seen rapid adoption across development tools and AI clients, and it now sits alongside other emerging standards aimed at agent interoperability. As protocols mature, maintainers commonly prune features that hinder scalability or add implementation burden, favoring designs that behave predictably under distributed load. A move toward statelessness echoes long-standing web architecture principles, where self-contained requests have proven easier to cache, route, and scale.

As with any protocol-level change, the timeline for ecosystem support will vary. Client libraries, server frameworks, and hosted MCP offerings each update on their own schedules, so a period of mixed compatibility is likely. Developers are advised to test against updated tooling, monitor the specification's changelog, and plan migrations deliberately rather than assuming immediate, uniform support across the tools they depend on.