MCP Goes Stateless: The Protocol Connecting AI to External Data Gets Easier to Scale

The Model Context Protocol's latest release candidate, published July 28, 2026, on the official MCP blog, switches server-side session handling to a stateless model and adds mandatory routing headers to the Streamable HTTP transport. Both changes target the same goal: making MCP servers easier to run behind load balancers at scale.
The core change concerns session IDs. Under the previous design, MCP servers kept session state in memory, which complicated horizontal scaling — the practice of adding more server instances to handle more load. The problem: any given request had to land on the specific server instance holding that session. The release candidate moves to a stateless approach on the server side, so requests can be spread across instances without requiring session affinity, the practice of pinning a client to one server. Arcade, a startup founded by Nate Barbettini, published a technical breakdown at arcade.dev/blog/mcp-going-stateless/ walking through the changes and their operational implications. (TechCrunch)
A second specification change requires the Streamable HTTP transport to include Mcp-Method and Mcp-Name headers on requests. The official MCP blog states these headers make traffic easier to operate, enabling intermediaries (like proxies and API gateways) and observability tooling to identify and route MCP traffic without parsing the request body. (MCP Blog)
TechCrunch, reporting on the update on July 20, 2026, describes MCP as "one of the basic building blocks of AI interoperability, giving AI models a secure way to access external data sources and services." The specification has been public since May, according to the same article, with the July 28 release candidate formalizing changes that had been under discussion in the interim. (TechCrunch)
Anthropic created MCP as an open standard for AI applications to connect to external tools and data. In December 2025, the company announced it was donating the protocol and establishing the Agentic AI Foundation as a neutral home for its governance. An official, community-driven MCP Registry was launched at the same time for discovering available MCP servers. (Anthropic)
The ecosystem around MCP has been expanding on several fronts. Reuters launched an MCP server on July 8, 2026, to deliver news content directly into AI workflows. On July 16, 2026, the Agentic AI Foundation introduced the MCPA, described as the first official certification for the Model Context Protocol. Arcade, which published the stateless transition walkthrough, raised $60 million in June 2026. (Reuters) (BusinessWire) (Agentic AI Foundation)
Anthropic's own engineering work has steadily pushed MCP into production use. A June 2025 engineering post described "Desktop Extensions," enabling one-click MCP server installation for Claude Desktop. A November 2025 post detailed code execution via MCP, explaining how the approach lets agents handle more tools while using fewer tokens and reducing context usage. Anthropic has also published support documentation for custom connectors using remote MCP and, in October 2025, added connectors for S&P Capital IQ and Daloopa as part of advancing Claude for Financial Services. In August 2025, Anthropic and Google Cloud co-presented a webinar on deploying multi-agent systems using MCP alongside the Agent-to-Agent (A2A) protocol with Claude on Vertex AI. (Anthropic Engineering) (Anthropic Engineering) (Anthropic Support) (Anthropic) (Anthropic Webinars)
The stateless transition addresses a genuine operational pain point. Anyone who has deployed stateful services behind a load balancer knows the friction: sticky sessions reduce flexibility, instance failure breaks active sessions, and autoscaling becomes a series of compromises. MCP servers exposed to high-volume AI inference pipelines face the same constraints. Removing server-side session state means a request can hit any healthy instance, which simplifies deployment topology and failure recovery. The header requirement complements this by giving network intermediaries enough metadata to route and monitor traffic without deep packet inspection.
The broader context here is that these two changes lower the operational cost of running MCP infrastructure at a moment when adoption is accelerating. Reuters building an MCP server, the Agentic AI Foundation introducing certification, and Arcade securing funding to build tooling around the protocol all point to MCP moving from specification toward production standard. The protocol's donation to a foundation and the launch of a community registry, both in December 2025, established the governance scaffolding. This release candidate does the quieter, less glamorous work of making the plumbing reliable enough that infrastructure teams will adopt it without caveats.
For engineers currently implementing MCP servers with session state, the release candidate is worth evaluating against existing deployment patterns. The stateless model shifts responsibility for any required state to the client or to an external store, which is the standard tradeoff. The header changes are lower-risk but require updates to transport-layer code. Both are backward-incompatible and will require coordinated client and server updates.


