MCP, the open standard for AI agents, drops its session handshake so it can run on plain Kubernetes

The Model Context Protocol, the open standard AI agents use to call tools and pull in context, just had its wire contract reworked to look like a normal web service. The 2026-07-28 specification release from the Agentic AI Foundation (AAIF) retires the protocol's session handshake, adds header-based routing, and ships a formal 12-month deprecation policy. Together, the three changes turn MCP into a workload class that conventional Kubernetes and load balancers can already carry.

"This is v2 in spirit, and the biggest change we've ever made to the protocol," David Soria Parra told VentureBeat. The new getting-started docs frame the change in three lines: "MCP is now stateless at the protocol layer. Any request can land on any instance. State moves on the wire."

What changed at the wire

Before 2026-07-28, an MCP client and server negotiated a session during the initialize/initialized exchange and exchanged a Mcp-Session-Id header on every subsequent request. That required sticky sessions in a load balancer, or a shared session store, to keep traffic coherent. SEP-2575 retires both: per-request _meta now carries the protocol version and client identity, and the new server/discover RPC is available but optional. "Any MCP server behind a plain round-robin load balancer with no shared session storage can serve the protocol correctly," Caitie McCaffrey told The Register.

The second change is header-based routing. The 2026-07-28 spec makes Mcp-Method and Mcp-Name mandatory on Streamable HTTP, so gateways, WAFs, and rate limiters can route and authorize on headers instead of parsing JSON bodies. That is the same pattern HTTP, gRPC, and modern API gateways already use, and the one enterprise platform teams have runbooks for.

Long-running agent flows also needed reshaping. Server-initiated elicitation/create, sampling/createMessage, and roots/list were stateful round trips that only worked over a live session. The release replaces them with Multi Round-Trip Requests, a request/response shape using resultType: input_required plus inputResponses, which works over the new stateless transport. The Tasks extension formalizes the pattern with tasks/get (poll) and a new tasks/update, so a task handle survives a pod restart and can be polled by any instance.

The enterprise lifecycle

The third change is governance. SEP-2596 introduces a formal Specification Feature Lifecycle and Deprecation Policy: at least 12 months between deprecation and removal. Roots, Sampling, Logging, and the legacy HTTP+SSE transport are now deprecated under that policy. Caitie McCaffrey, in comments to The Register, called durable task handles and the new lifecycle the "operational payoff" of the new architecture.

The 2026-07-28 release requires authorization servers to return the iss parameter per RFC 9207 and clients to validate it before redeeming a code, closing the OAuth mix-up attack class Dynamic Client Registration was prone to. Dynamic Client Registration is formally deprecated in favor of Client ID Metadata Documents, and client credentials are bound to the issuing authorization server, so a credential stolen from one tenant cannot be replayed against another.

Trade-offs, and the migration window

Per-request payloads grow, and any tooling that assumed in-process state has to be rewritten. GitHub maintainer Den Delimarsky, in comments to VentureBeat, framed the design as a deliberate transfer of state responsibility to developers. The release also drops out-of-band server logging; an adoption scrape turned up essentially no users. Existing stateful deployments get the 12-month migration window under SEP-2596.

Close to half a billion SDKs downloaded per month. The AAIF release post reports that figure across the Tier 1 languages, with the TypeScript and Python SDKs each past one billion total downloads. AAIF executive director Mazin Gilbert, in comments to VentureBeat, said member companies are deploying "tens of thousands of agents" that were blocked on the old architecture; that is a member-reported characterization, not a measured population. Tier 1 SDKs (TypeScript, Python, Go, and C#) shipped the new spec on day 0.

The next checkpoint is the December 2026 deprecation cutoff for the legacy transport and Roots/Sampling/Logging, with the formal 12-month window under SEP-2596 closing in mid-2027.