Model Context Protocol Becomes Stateless and Launches Race for Edge and Serverless Agent Deployment

On July 28, Anthropic published the new revision of the Model Context Protocol, version 2026-07-28, featuring an engineering decision that alters how AI agents will be deployed in production. The MCP transitions from being a protocol with a persistent session between client and server to operating in a request-response model, without an initialization handshake and without state at the protocol level.

The revision removes both the initialize handshake and the session at the protocol level. Each request carries the version of the MCP, the client identification, and the negotiated capabilities. MCP servers now run in ephemeral runtime: any instance can handle any request, without sticky routing or shared session storage. Multi Round-Trip Requests, header-based routing, cacheable lists, and hardening of authorization routines with OAuth 2.1 and OIDC complete the technical package. The Tier 1 SDKs, maintained by Anthropic, Microsoft, and AWS, have already been updated.

The practical consequence is felt in infrastructure costs. A stateful MCP server required the same instance to handle all calls from a client during a session, which forced sticky routing on load balancers and distributed caching to maintain context. A stateless MCP server runs like Lambda or Cloud Function, scales to zero, and does not incur costs for idle connections. The model resembles that of REST in the early 2010s. Operators managing fleets of agents in Kubernetes with session affinity are beginning to view the architecture as legacy.

Extensions and the AWS Hook

The revision formalizes a versioned extension framework. Two official extensions are launching on day one: Apps, contributed by Anthropic to standardize the presentation of interactive content via MCP servers; and Tasks, contributed by AWS for long-running agents with reentrancy guarantees. This is the institutional design that MCP previously lacked and is the hook that AWS used to position Amazon Bedrock AgentCore as the reference platform. The AgentCore announcement comes bundled with the promise of running MCP servers in serverless runtime with session isolation starting that same afternoon.

Today, the MCP is the only protocol with cross-organizational adoption among Anthropic, OpenAI, Microsoft, Google, and AWS. Anthropic reported 400 million monthly downloads of the SDKs, four times the volume from earlier this year. Transitioning to stateless reduces the marginal cost of operating an MCP server in production and dismantles the scalability barrier that prevented smaller startups from exposing their own tools as public servers. This change opens the door for a genuine MCP server marketplace, rather than just example repositories on GitHub.

B2B Reading by Geography

American banks like JPMorgan and Goldman Sachs, which are already running internal agents anchored in MCP for research and onboarding, now have the opportunity to expose regulatory tools at the edge, closer to clients in Europe and Asia, without replicating entire backends. Global consultancies with engineering arms such as Capgemini, Accenture, and TCS see lower operational overhead for the customized agents they have been selling case by case to manufacturing clients in Germany and Japan, where data residency requirements complicate centralized runtime.

In Brazil, engineering teams from fintechs and consultancies like CI&T, Falconi, and the AI arm of B3 had resisted adopting MCP in production due to the challenges of maintaining sessions. This objection no longer exists. The conversation that is currently centered around proprietary MCP gateways and internal brokers is likely to shift towards serverless orchestration using the same standard protocol that the rest of the market employs.

What the Change Does Not Resolve

Transitioning to stateless resolves infrastructure cost issues. It does not resolve security around tool execution. The standardization of OAuth 2.1 and OIDC in the protocol is a formal advancement, but the real risks remain on the MCP server side that exposes tools to the model: prompt injection against tool descriptions, abuse of delegated permissions, and resource enumeration via error responses continue to be open vulnerabilities. Cases such as the reported vulnerable MCP inspector from May of this year are still relevant. The July 28 revision shifts the agent scaling problem to an agent governance problem, which is where the CISO will now focus.

Engineering came before policy. The July 28 revision represents the most significant change to the MCP since the version 2024-11-05, and timing is critical. It arrives as Anthropic faces competition from OpenAI in the same agent landscape, with AWS seeking to avoid the mistake of allowing LangChain to become the standard's steed. Those who bet on the MCP as a neutral standard now hold the technical justification for their investment. Conversely, those who invested in proprietary tools find a narrower window to justify their architecture.