MCP goes stateless: build your agent integrations on the standard

The protocol that connects AI agents to your tools and data just had its biggest revision since launch — and the headline change is that it's getting simpler. The Model Context Protocol's 2026-07-28 release candidate makes MCP stateless at the protocol layer, which sounds like a plumbing detail and is actually a signal worth acting on: the standard your agent integrations will speak is stabilizing enough to build on.

What actually happened

Per the MCP spec blog, the release candidate was locked May 21 with the final spec landing July 28, 2026, after a ten-week window for SDK maintainers to test against real workloads. The big change: MCP drops the initialize/initialized handshake and the Mcp-Session-Id header that used to pin a client to one specific server instance. Servers that needed sticky sessions, a shared session store, and deep packet inspection at the gateway can now sit behind a plain round-robin load balancer. When an app genuinely needs state, tools return explicit handles that the model passes back as arguments — state becomes visible and intentional instead of hidden in the transport.

The rest of the release reads like a protocol growing up: authorization hardening (clients validate the iss parameter per RFC 9207), a formal extensions framework with the first two official extensions — MCP Apps (server-rendered UIs in sandboxed iframes) and Tasks — full JSON Schema 2020-12 for tool definitions, and header-based routing plus client-side caching of tool/resource lists. Older features like Roots, Sampling, and Logging are deprecated under a new lifecycle policy that guarantees at least twelve months before anything is removed.

Why it matters for your business

For most operators, MCP is invisible — it's how the AI you use reaches your calendar, your inventory, your customer records. But "invisible" is exactly why the standardizing matters. Two years ago, every AI-to-tool connection was a bespoke integration that broke when a vendor shipped an update. A stable, stateless standard means the connective tissue between your systems and whatever model you're using this quarter stops being a one-off you re-pay for every time you switch tools.

The stateless redesign is also a portability win you can feel in the bill: stateless servers are cheaper to run and far easier to scale, which is the difference between an integration you host yourself and one you rent from whoever locked you in. If you're commissioning custom AI work in 2026, the question to ask your builder is simple — are we building on MCP, or on some vendor's proprietary connector that only works until they change their mind? Build for the standard, and the plumbing you pay for once keeps working when the models underneath it don't.

Sources: Model Context Protocol — 2026-07-28 Release Candidate.