MCP Expands: Stateless Agent-Servers Finally Scale Without Limits

MCP 2026-07-28 drops session IDs and the initialize handshake. Every request carries the protocol version and client info itself. This allows agent servers to run behind standard round-robin load balancing and on serverless, without sticky-session workarounds.

Key takeaways

Related:Model Context Protocol under the Linux Foundation  /  AWS removes agent infrastructure-with a catch

Why Session State Slowed Agent Infrastructure

Until the July 28, 2026 spec, remote MCP relied on long-lived sessions. Load balancers required sticky affinity. Serverless functions, which go dormant after each request, were a poor fit. Teams built Redis session stores or kept containers warm just to ensure a tool call hit the same pod.

The new spec flips the model. Every JSON-RPC request is self-descriptive. Protocol version, client identity, and capabilities travel in _meta. If a client wants to know server capabilities upfront, it calls server/discover. This isn’t mandatory. Any instance behind a simple round-robin load balancer can handle any request.

What Platform Teams Must Retool Now

Three areas demand attention; the rest is migration.

First, transport. Session IDs in headers are dead. If you need state across tool calls, the model returns an explicit handle as an argument. The handle is visible-clunkier than hidden session state, but far more debuggable.

Second, multi-round-trip requests. Previously, the server held a stream open when it needed confirmation or a parameter mid-tool-call. MRTR now returns resultType: “input_required”. The client responds and retries the call with inputResponses. Elicitations and sampling no longer require persistent streams.

Third, authentication. Issuer validation per RFC 9207 is mandatory. Dynamic Client Registration is deprecated in favor of Client ID Metadata Documents. CLI clients using localhost redirects must set application_type correctly-or OAuth will still fail on redirect URIs.

Header Routing and Cacheable Lists

Streamable HTTP requests must set Mcp-Method and Mcp-Name. Gateways can meter and authorize tools without parsing the body. This sounds minor but transforms daily operations: per-tool rate limits, tenant-specific allowlists, audit logs with readable method names.

Responses from tools/list, prompts/list, and resources/list include ttlMs and cacheScope. Clients cache tool catalogs. Prompt caches remain stable across reconnects. Pulling the tool list every 30 seconds wastes latency and token budget.

Tasks, Deprecations, 12-Month Window

Tasks are moving from the core to the extension io.modelcontextprotocol/tasks, with poll-based tasks/get and tasks/update. Long-running agent processes are now specified, no longer improvised. Roots, sampling, and logging are deprecated but will remain usable for at least twelve months. New implementations should no longer adopt them.

The open spec provides the framework. Anthropic reports over 950 MCP servers in the Connectors directory for Claude and is building MCP apps for interactive interfaces, centrally managed enterprise authentication, and observability for connectors. For platform teams, this brings identity, permissions, telemetry, and the release of UI components under the same operational checks as transport and scaling.

SDKs for TypeScript, Python, Go, and C# support the spec. AWS is promoting the stateless core for Amazon Bedrock AgentCore. Cloudflare Agents SDK and Microsoft Foundry are signaling day-zero support. This isn’t just a marketing footnote: it means MCP servers will soon run like normal HTTP workloads, not as special cases with session stickiness.

An underrated side effect: list caches stabilize prompt caches. If tool catalogs arrive in a different order with every reconnect, prompt prefix caches fragment. Deterministic ordering plus TTL isn’t just a convenience feature-it saves tokens and makes agent runs cheaper.

Teams running MCP behind API gateways can attach method headers to existing policy engines. Allowlists per tenant, denials for destructive tools in production, burst limits per MCP name. This was messy and slow with pure body parsing. With headers, it becomes operationally standard.

A migration sequence that works in practice: first update clients and SDKs to 2026-07-28, then servers without session assumptions, then harden authentication, then phase out old transports. Breaking auth first just generates support tickets without scale benefits.

What to Check on Monday

Inventory your MCP servers that assume sessions. Verify whether your ingress enforces sticky sessions. Switch authentication to issuer validation and CIMD paths. Plan the breaking migration of SDKs like any other protocol upgrade: staging, canary, then production. If you’re only running MCP as a local demo, you can wait. But if you’re deploying agent tools in production, 2026-07-28 finally brings the HTTP semantics the rest of your stack already uses.

Frequently Asked Questions

Image source: AI-generated (July 2026)