MCP goes stateless, sparking debate over cloud dependency

The Model Context Protocol (MCP) has undergone its most significant overhaul to date, transitioning from a stateful, session-based design to a fully stateless architecture that dramatically enhances scalability and operational simplicity. This shift, championed by contributors from Anthropic and supported by industry leaders like AWS and Cloudflare, removes the need for sticky load balancers and shared session stores, allowing any MCP request to be routed to any server instance seamlessly. As Anthropic’s David Soria Parra noted, this change aligns MCP with cloud-native service patterns, enabling deployments on Kubernetes and standard HTTP infrastructure, which was previously an operational bottleneck for large-scale AI agent workflows. (Insights: 1, 5, 11, 17, 20, 24, 25, 32, 37, 42, 46, 52, 56, 60, 63)

To complement the stateless core, MCP introduces explicit handles—such as basket_id—that replace hidden session state and become visible to AI models, enabling more powerful agentic workflows where models can reason about and pass state across multi-step tool chains. This architectural redesign shifts responsibility for state management from the protocol layer to clients and cloud gateways, a move that simplifies the MCP spec but centralizes persistent state handling in infrastructure-heavy platforms like AWS’s AgentCore Gateway and Snowflake Cortex. While this offloading eases protocol complexity, it also raises operational costs and dependency considerations for enterprises and smaller developers alike. (Insights: 14, 18, 29, 30, 31, 35)

The update streamlines MCP by deprecating niche or overly complex features such as sampling, roots, and verbose logging, replacing them with modern standards like JSON Schema 2020-12 and recommending logging via OpenTelemetry or stderr. This pruning, alongside the introduction of a formal extensions framework, allows for domain-specific capabilities—like interactive server-rendered interfaces and long-running asynchronous tasks—to be developed and tested independently before core integration. Such modularity supports diverse enterprise use cases while maintaining protocol lean-ness and future-proofing. (Insights: 3, 4, 23, 28, 40, 41, 53, 64)

Security and governance have been significantly hardened in the new MCP specification, which aligns authorization mechanisms with OAuth 2.1 and OpenID Connect standards and introduces a formal 12-month deprecation policy to provide enterprises with predictable migration timelines. Enhancements like SEP 2468 mandate issuer validation to prevent OAuth Mixup Attacks, while new enterprise-managed authorization extensions developed with Octopus enforce corporate login requirements, ensuring secure data handling in large-scale deployments. These improvements reflect a maturation of MCP into an enterprise-grade protocol ready for robust, secure AI agent orchestration. (Insights: 8, 22, 26, 27, 36, 44, 48, 49, 54, 57, 58)

Despite widespread enthusiasm for MCP adoption, many companies have implemented it as a superficial additional layer that merely mimics existing APIs rather than embracing its original design intent. Moe Katib critiques this trend, emphasizing that MCP was initially created for direct developer use on local machines to seamlessly connect with existing resources, but enterprise-scale deployments have diluted this vision, leading to inefficiencies and missed opportunities for deeper integration.

While MCP has eased the technical challenge of integrating AI with legacy systems, companies like Musinsa Tech highlight persistent difficulties in aligning AI behavior with developer intent and maintaining output quality. Their solution involved granular input validation and strict intent identification at the MCP server level, underscoring the ongoing struggle to govern AI’s inherently non-deterministic nature and ensure it operates within carefully defined boundaries, particularly when applying the principle of least privilege to protect sensitive data.

Concerns about MCP’s technical maturity remain, especially regarding security, authentication, and governance standards necessary for broad enterprise adoption. The protocol’s recent removal of session state management simplifies the specification but shifts the complexity and cost of state persistence to cloud gateways like AWS, Agent Core, and Snowflake Cortex. This architectural pivot creates a dependency on large infrastructure providers, complicating production workflows for smaller developers and firms who now face higher operational costs despite a seemingly simpler protocol.

The Model Context Protocol (MCP) has rapidly evolved into a foundational infrastructure enabling seamless interoperability across diverse enterprise AI applications. By early 2026, MCP was deeply embedded within Microsoft’s ecosystem, powering integrations from SQL Server 2025 to Dynamics 365 Sales, where it connects AI agents to both internal and external data sources such as ZoomInfo and Dun & Bradstreet. This widespread adoption, reflected in 97 million monthly SDK downloads and over 10,000 public servers, underscores MCP’s role as a structural assumption in enterprise AI rather than a nascent standard, facilitating unified workflows that blur traditional boundaries between CRM, ERP, and commerce systems.

AWS’s practical implementation of MCP showcases how agent-to-agent communication can enhance existing business workflows without costly rebuilds. By integrating the AWS Partner Central Agent via MCP, organizations enable natural language exchanges where custom orchestrators express intents while the Partner Central Agent manages business rules, validations, and approvals. This architecture not only streamlines co-sell opportunity management directly within established workflows but also improves deal velocity and revenue growth by eliminating disruptive context-switching for sales teams.

In the business travel sector, MCP is driving a transformative shift from isolated AI assistants to connected AI ecosystems capable of executing complex, multi-departmental workflows. Enterprises leveraging MCP-enabled AI copilots can securely interact with travel bookings, expense policies, and financial systems, automating tasks from itinerary management to expense reconciliation while enforcing role-based permissions and audit trails. This standardization reduces the need for costly custom integrations and enhances operational efficiency, as AI assistants become trusted partners across travel, finance, procurement, and HR functions.

Vidmob’s Signal 78 platform exemplifies MCP’s growing impact in Martech by aggregating business context through two-way connections with clients’ software systems. This integration empowers AI assistants to operate marketing stacks via natural language, eliminating the traditional reliance on developers for API integrations. While accelerating adoption among major Martech players like 6 Sense, Sprinklr, and Adobe, this shift also raises governance imperatives around role-based permissions, human approvals, and audit trails to ensure secure and accountable AI-driven actions within enterprise environments.