Anthropic's MCP goes stateless in its biggest overhaul yet, answering dev calls for scale

The Model Context Protocol (MCP) just received its most significant architectural change since it launched. Lead maintainers David Soria Parra and Den Delimarsky at Anthropic announced that the protocol’s core is now stateless, ditching the old bidirectional, session-dependent model for a straightforward request/response pattern. This isn’t a minor tweak. The shift directly addresses what Soria Parra called “one of the most highly-requested features from developers” who were hitting reliability and scalability ceilings with the original design. No longer tying a request to a specific server instance’s session fundamentally changes how AI systems can connect to external tools and data. It opens the door to much more distributed, fault-tolerant architectures.

Beyond the stateless core, the update bundles several other substantial improvements. The spec now supports Multi Round-Trip Requests and header-based routing, which gives developers finer-grained control over how complex interactions are managed. List results are now cacheable, a practical change that should reduce latency and overhead for common operations. The team also hardened authorization and introduced a formal extensions framework. This framework is particularly interesting because it creates a structured way for the community to extend MCP’s capabilities without forking the core protocol. Updated Tier 1 SDKs round out the release, ensuring the most popular languages get first-class support for these new features immediately.

Soria Parra described this as “MCP’s most important since remote MCP first launched over a year ago.” That’s not hyperbole. The original stateful model was a known bottleneck. It made horizontal scaling difficult and created single points of failure. For an open standard that’s meant to be the universal bridge between AI models and the outside world, that was a growing liability. This move shows Anthropic is listening to the engineers actually deploying MCP in production, not just designing in a vacuum. The stateless protocol core positions MCP more like HTTP—simple, scalable, and resilient by default.

What’s left unsaid is how this will play against OpenAI’s function calling and Google’s tool use APIs. MCP’s value proposition has always been its openness and interoperability. A more robust, scalable protocol makes that pitch stronger. But protocol updates only matter if adoption follows. The extensions framework is a smart hedge here—it lets the ecosystem evolve without a central bottleneck. The next test isn’t the spec. It’s whether the tooling and infrastructure around MCP can make this new stateless reality as easy to deploy as the blog post makes it sound.