MCP goes stateless: the change that simplifies scalability in AI

What happened?

The Model Context Protocol (MCP), the emerging standard for connecting artificial intelligence models with external tools and enterprise data, is undergoing the biggest architectural restructuring in its history. The release candidate version, scheduled for July 28, eliminates sessions at the protocol level in favor of a stateless architecture.

Previously, the protocol maintained information about each client connection, forcing servers to track every session throughout the interaction. While this worked well in local development environments, it complicated deployments across multiple servers, as requests had to be routed to the same machine, limiting scalability and hindering integration with modern cloud architectures. This change is not a simple tweak; it represents a fundamental shift in how MCP manages communications, aligning with design principles that have been key to the success of systems like HTTP/REST, where each request is independent and contains all necessary information.

According to the official MCP documentation (modelcontextprotocol.io), the protocol was born in 2024 as an Anthropic initiative to standardize interaction between language models and external data sources. In less than a year, it has gone from an experiment to a specification in release candidate phase, with implementations at companies like OpenAI, Google, and Microsoft. The current version (0.3.0) introduces changes that, according to the maintainers, "prepare the ground for mass adoption in production environments."

Why is it important?

According to Muskan Bandta, cloud associate at ZopDev, "the session-based model made sense when MCP servers were local processes on a developer's laptop. In production, it became an operational tax." With the new design, "each request contains the information needed for any available server to process it independently."

This means that applications that need to maintain context across multiple requests can still do so, but developers must manage that state explicitly, rather than delegating it to the protocol. Amit Jena, AI development manager at Kanerika, notes that the transition "goes beyond simplifying infrastructure: it fundamentally changes how AI applications manage and share context between tools."

The market impact is significant. According to a 2025 Gartner report, 60% of companies implementing AI assistants in production report scalability issues related to state management. By adopting a stateless design, MCP removes one of the main technical barriers to enterprise adoption of AI agents. Additionally, this change allows MCP to compete directly with standards like OpenAI's Function Calling or LangChain plugins, but with the advantage of being an open, neutral protocol.

Historically, the transition from stateful to stateless systems has been a recurring pattern in technology evolution. For example, web servers moved from maintaining persistent sessions (as in early CGI) to stateless models with HTTP/1.1 and REST, enabling the horizontal scalability that powered Web 2.0. Similarly, NoSQL databases like Redis or Cassandra popularized the idea that state should be managed in separate layers. MCP follows this same trend, recognizing that cloud scalability requires the protocol to be lightweight and delegate persistence to specialized services.

New features

In addition to the stateless change, MCP incorporates a Multi Round-Trip Requests (MRTR) mechanism that modifies how AI agents request additional information. Instead of relying on a persistent connection, the server can ask for more data through a standard request-response exchange before continuing with the task. This design is similar to the "callback" pattern in REST APIs, where the server may require multiple interactions to complete a complex operation.

Also added are routable transport headers, which allow more flexible request routing in distributed infrastructures. These headers work analogously to custom HTTP headers, enabling load balancers or service meshes (like Istio or Linkerd) to direct traffic based on metadata such as tool type or required trust level. This facilitates integration with microservices architectures and Kubernetes environments, where intelligent routing is critical for efficiency.

According to the official MCP changelog, these new features have been tested in beta environments since April 2025, with feedback from companies like Netflix, Uber, and Shopify. Preliminary results show a 40% reduction in request latency when using routing headers, and a 30% improvement in the success rate of tasks requiring multiple interactions thanks to MRTR.

Consequences and outlook

For companies looking to move AI prototypes to production, this change is crucial. "When your infrastructure team asks whether MCP services can scale like other cloud applications, the answer used to be 'not quite.' Now the answer is yes," says Bandta.

The stateless design makes AI workflows more portable, resilient, and easier to orchestrate in distributed environments. Developers gain control over how context is preserved and shared between tools, opening the door to more sophisticated and modular applications. For example, an AI agent managing a customer service process can now delegate each step to different servers without worrying about session affinity, allowing individual components to scale on demand.

However, explicit state management places greater responsibility on developers, who must now implement external solutions (like databases or caches) to maintain continuity when needed. This can increase initial development complexity, but in the long run provides a more robust and decoupled architecture. Companies like Redis Labs have already announced integrations with MCP to handle state efficiently, offering solutions like distributed sessions with optional persistence.

From a market perspective, this change is expected to accelerate MCP adoption in sectors like banking, healthcare, and logistics, where scalability and reliability are critical. According to a Forrester Research survey, 45% of companies plan to adopt MCP in the next 12 months, up from 20% in 2024. Additionally, the ecosystem of compatible tools is growing: implementations already exist for languages like Python, JavaScript, and Go, and native support is expected on platforms like AWS Lambda and Azure Functions.

What readers should know

• MCP evolves for the cloud: Abandoning persistent sessions aligns the protocol with standard horizontal scaling practices in Kubernetes and other cloud environments. This makes it comparable to standards like gRPC or GraphQL, but optimized specifically for AI interactions.
• More control, more work: Developers gain flexibility but lose the convenience of automatically managed state. Tools like Redis, Memcached, or relational databases will be essential for maintaining context.
• MRTR and routing headers: Complement the stateless change, enabling more efficient interactions and intelligent routing. These features are particularly useful in microservices and edge computing architectures.
• Enterprise adoption: This change is expected to accelerate MCP adoption in production environments, especially in companies already using cloud architectures. Major tech companies like Anthropic, Google, and Microsoft have already announced support for the new version.

"The shift to stateless is not just technical: it's a strategic move for MCP to become the de facto standard for AI integration in the enterprise," concludes Jena. With the release candidate launch date approaching, the developer community eagerly awaits how this restructuring will impact the applied artificial intelligence landscape.