AWS has outlined how the latest Model Context Protocol (MCP) specification changes remote MCP server deployments by removing protocol-level sessions and allowing requests to reach any available server instance. The change eliminates the protocol requirement for sticky sessions and shared session stores, simplifying horizontal scaling while shifting state management and other responsibilities to surrounding infrastructure.
The updated MCP specification removes the initialize and initialized handshake and the Mcp-Session-Id header. Requests can therefore be routed independently to any server instance behind a conventional load balancer. The specification also introduces an optional server/discover operation for clients that need server capabilities before making tool calls.
AWS diagram showing how MCP changes map to the Well-Architected Agentic AI Lens. Source: AWS Architecture Blog.
For AWS deployments, this can eliminate infrastructure used specifically to maintain MCP protocol sessions. AWS Architecture Blog authors Anand Komandooru, Steven DeVries, and Haleh Najafzadeh describe replacing session-affine routing with conventional request routing and removing session storage used solely for MCP protocol state. They also identify AWS Lambda as a deployment option that fits the request-response model because the protocol no longer requires persistent session connections.
The distinction between protocol and application state has also emerged in community discussion. Michael Madsen, writing about the specification on LinkedIn, summarized the change as:
The protocol is stateless. Your application doesn't have to be.
MRTR replaces server-initiated requests that previously required held-open streams, allowing multi-step interactions through `input_required` responses and subsequent requests. New `Mcp-Method` and `Mcp-Name` headers enable gateway routing and throttling, while W3C Trace Context supports distributed tracing. ttlMs and cacheScope provide caching controls.
AWS maps these changes to its Well-Architected guidance for agentic AI, covering monitoring, tracing, security, and tool integration. Stream resumability has also been removed, so clients may need to retry interrupted operations, increasing the importance of idempotency for tool calls that produce side effects.
Early implementation work shows that existing infrastructure still requires a transition path. Apify's MCP server project is implementing stateless support alongside its existing sessionful server, with routing and conformance tests covering both protocol versions.
Migration therefore remains relevant for deployments supporting earlier MCP clients. AWS recommends tracking protocol versions at the gateway and retaining session infrastructure until legacy traffic has been eliminated. The MCP project has also established a feature lifecycle policy providing a defined migration period for deprecated capabilities.