Skip to content

MCP drops session pinning so agent servers can scale normally

Original: Scaling AI Agent Infrastructure with the MCP Stateless updates View original →

Read in other languages: 한국어日本語
LLM Aug 6, 2026 By Insights AI 2 min read 1 views Source

A core piece of agent infrastructure is moving away from session pinning. In an August 5, 2026 post, Google said the 2026-07-28 Model Context Protocol specification release candidate removes transport-level session management. The change matters because MCP started as a clean local integration protocol, but stateful transport became a deployment tax once teams tried to run remote MCP servers across Kubernetes, load balancers, and serverless platforms.

The old HTTP model required an initialization flow where the server returned a Mcp-Session-Id. Every later tool call or resource query had to return to the same container or pod holding that in-memory session. Put three pods behind a standard round-robin load balancer, and a second request could land on a different pod and return 400 Session Not Found. Teams then had to add sticky sessions, shared Redis session stores, or gateway packet inspection.

The new model makes each request self-contained. Protocol version, client info, and capabilities now travel inline in a request _meta field rather than being negotiated once at connection setup. The initialize / initialized handshake and logical Mcp-Session-Id header are removed. Google frames this as the largest MCP spec change since launch because it lets ordinary HTTP load-balanced infrastructure handle any request on any healthy server instance.

The operational details are concrete. New HTTP headers such as Mcp-Protocol-Version, Mcp-Method, and Mcp-Name let proxies and gateways route, rate-limit, and audit calls without reading the request body. Multi Round-Trip Requests handle user confirmations or server-to-client prompts without holding open a long-lived SSE connection. Security changes include issuer verification, resource indicators, and JSON Schema 2020-12 support for tool inputs.

The migration path is already visible. Google says the four Tier-1 SDKs for TypeScript, Python, Go, and C# have beta releases supporting the 2026-07-28 spec. It also says production servers such as GitHub MCP Server have upgraded and removed Redis session storage. Roots, Sampling, and Logging enter a formal deprecation policy with a minimum 12-month transition window. For teams building enterprise agent systems, MCP is shifting from a local adapter pattern toward infrastructure that can survive rolling deploys, autoscaling, and plain round-robin routing.

Share: Long

Related Articles