MCP 2026-07-28: the stateless core and why it changes servers

The short answer

The Model Context Protocol 2026-07-28 spec is one of the most significant MCP releases to date, and its headline is a stateless core.

MCP moves from a bidirectional, stateful protocol to a request/response model. That single change ripples through everything: servers no longer need to hold a persistent session, which makes them deployable on serverless, edge, and standard load-balanced HTTP infrastructure instead of only long-lived processes.

What actually changed

• Stateless core. The protocol drops the concept of a maintained session between client and server. Each request is self-contained and answered on its own — closer to how HTTP works than to a stateful duplex connection.
• Extensions framework. The same release introduces standardized extensions including MCP Apps and Tasks, plus hardened auth.
• Enterprise leverage. Anthropic frames the value as letting teams "build and connect agentic experiences at enterprise scale" — an explicit nod to how much simpler stateless deployments are to operate.

Why stateless matters for server builders

If you build MCP servers, the stateful model was the awkward part: you had to keep a long-running process alive, manage session lifecycle, and handle reconnects. Stateless removes that burden.

A stateless server is just an HTTP endpoint that answers tool calls. That means:

• It runs on serverless functions (scale to zero, no idle process cost).
• It runs on edge and load-balanced HTTP infrastructure with no session-affinity concerns.
• It's trivially replicable and horizontally scalable — any instance can answer any request.

This is the infrastructure shift that lets AI agents connect to tools the same way web apps connect to APIs.

Context: the protocol is already everywhere

The stateless change lands on a protocol that's already the industry standard for connecting agents to tools and data. Anthropic reports MCP surpassed 400M monthly SDK downloads — a 4x increase this year — with Microsoft, Google, and a growing roster of vendors shipping MCP compatibility.

So the practical takeaway isn't "learn a niche protocol." It's: the default way agents reach external systems just got a lot cheaper and easier to run.

The parallel move: graph engineering

The stateless shift pairs with a broader change in how agentic systems are built. The same period has seen a move from loop engineering (everything inside one context window) toward graph engineering — where subagents act as nodes, an orchestrator coordinates them, and shared state lives outside any single window.

MCP's stateless core supports this direction: if servers are stateless endpoints, graphical, parallel, fan-out architectures become much easier to wire up.

Takeaway

The 2026-07-28 spec is a quiet but load-bearing change: MCP servers no longer have to be stateful processes. If you've been putting off building an MCP server because persistent session management felt heavy, that reason is gone. A stateless tool endpoint is now a modest HTTP service you can host almost anywhere — the same model that made modern web APIs trivial to operate.

---

Primary source: Bringing MCP 2026-07-28 to Claude (Claude/Anthropic). Stateless core, extensions, auth, and the 400M download figure verified against that post on 2026-08-08. The "graph engineering" framing derives from community commentary, flagged as interpretation, not an Anthropic claim.