In early 2025, when Anthropic introduced MCP (Model Context Protocol), most developers filed it as another “USB-C for AI” — a standard for connecting models to tools. Sound idea, nowhere near production.
Eighteen months later the story has completely changed. TypeScript and Python SDK downloads have each passed 1 billion cumulative, monthly downloads approach 400 million, and more than 950 MCP servers run in the Claude app store. MCP has become the de facto industry standard for AI tool interconnection.
And on July 28, Anthropic shipped the largest update in MCP’s history — the 2026-07-28 specification. Not a feature drop: a ground-up restructuring of the protocol architecture.
What Actually Changed
Cut one: the core goes stateless
The biggest change, and the most requested. Old MCP required a bidirectional stateful connection between client and server: the initialize/initialized handshake, then a persistent Mcp-Session-Id carried through the whole lifecycle. Your MCP server had to hold long-lived connections — which made deploying on Cloudflare Workers or AWS Lambda an exercise in frustration.
The new spec removes all of it. Every request is self-describing — protocol version, client info, and capability declarations ride along in the _meta field. Need capability negotiation? There’s an optional server/discover RPC. Don’t need it? Just send the request.
What that means: your MCP server now deploys behind any plain round-robin load balancer, and any instance can serve any request. Serverless, edge compute, horizontal scaling — the entire cloud-native toolchain suddenly fits MCP.
Cut two: MRTR — good riddance to long connections
Stateful protocols have a practical awkwardness: mid-execution, a server may need to ask the client for data (confirming parameters with the user, say). That used to require an open bidirectional stream. The new multi-round trip requests (MRTR) let the server return resultType: "input_required" with its questions; the client answers and re-invokes with inputResponses.
An elegant side effect: all streaming HTTP requests must now carry Mcp-Method and Mcp-Name headers. Your API gateway, rate limiter, and WAF can route and bill without parsing JSON bodies. From an operations standpoint, that is a one-sided upgrade.
Cut three: Extensions — from bolted-on features to a framework
Extensions are now first-class: reverse-DNS naming, independent versioning, delegable maintainers. Three official extensions launch with the spec:
-
MCP Apps: servers can render interactive UI directly inside the Claude chat interface, within a security-sandboxed iframe. Ask for revenue data and get a live dashboard; operate a cloud server and get a control panel in the conversation. The era of tab-switching is ending.
-
Tasks: the official async task framework. Long-running operations — analyzing 100GB of logs, rendering 3D video — now have a standard polling-plus-notification mechanism.
-
EMA (enterprise managed auth): centralized MCP server permissions through identity providers like Microsoft Entra and Okta. The administrator authorizes once; users inherit automatically — fully zero-touch.
Cut four: auth finally grows up
Auth integration consumed more developer hours than anything else this past year. The new spec does three things: aligns fully with production-grade OAuth 2.0 and OIDC deployments, enforces RFC 9207 issuer identification (closing the authorization server mix-up vulnerability), and formally deprecates DCR in favor of Client Metadata Documents (CIMD). Veterans of the old spec’s localhost redirect_uri errors will recognize how much pain just died.
Cut five: caching gets deterministic
tools/list and prompts/list responses now carry ttlMs and cacheScope with guaranteed deterministic ordering. Client-side caching policy stops being guesswork.
MCP’s Three-Year Arc, Framed
| Phase | Time | Core job | Markers |
|---|---|---|---|
| 1. Prove it exists | 2024–2025 | Get developers to accept "AI needs a standard protocol" | MCP 2024 release, first SDKs |
| 2. Make flows work | 2025 H1 | Basic feature set, mainstream scenarios | MCP 2025-03-26, HTTP+SSE transport, Streamable HTTP |
| 3. Land at scale | 2025 H2 | Enterprise deployment pain points | 400M monthly downloads, 950+ servers, Figma/Intuit/Netlify onboard |
| 4. Become infrastructure | 2026-07 | Architecture restructure for cloud-native | Stateless core, Extensions framework, enterprise auth |
This is the standard path from “works” to “usable” to “infrastructure.” The HTTP analogy is apt: early HTTP was also a simple text protocol, and it was precisely the stateless design that let it scale to the whole internet.
Where MCP Goes Next
Endpoint 1: MCP Tunnels (research preview)
Many enterprises cannot expose databases or ERPs to the public internet under any circumstances. MCP Tunnels connect Claude to internal networks with no public IPs, no inbound firewall rules, no IP allowlists — one encrypted tunnel through the corporate firewall. If it ships for real, this becomes the wedge that carries AI into finance, healthcare, and government.
Endpoint 2: the MCP Apps chain reaction
Rendered interactive UI turns Claude from a “chat window” into an “application runtime” — any MCP server can become an app. The end state: AI is no longer the user of tools but the platform for them. Figma, Intuit, and Netlify are already in; the next breakout is an open question.
Endpoint 3: the competitive picture
Google has A2A; OpenAI has GPT Actions. But MCP’s developer-ecosystem lead is enormous — 950+ servers, 400 million monthly downloads, mainstream SaaS vendors publicly committed. This architectural upgrade makes that lead close to unassailable: once a protocol becomes infrastructure, challengers face not an engineering gap but a network effect.
What Developers Should Do Now
-
Upgrade your SDKs. The four official SDKs — TypeScript, Python, Go, C# — all support the 2026-07-28 spec, with detailed migration guides. If you run MCP servers in production, the time is now.
-
Rethink deployment. The stateless core means Serverless is a first-class target. If your MCP server still runs a resident process, moving to Lambda or Cloudflare Workers cuts both cost and operational complexity.
-
Track the deprecation list. Roots, Sampling, Logging, and the old HTTP+SSE transport are deprecated with a 12-month grace period. Keep them out of new projects.
-
Experiment with MCP Apps. If you build user-facing AI tools, in-Claude interactive UI is a product-experience differentiator users can feel immediately.
MCP covered in eighteen months what HTTP needed years to travel from RFC to ubiquity. The 2026-07-28 version invented no new concepts — it did something harder: turned a protocol that worked in the lab into infrastructure that runs stable in production.
The genuinely interesting next step is not MCP itself. It is what gets built on top of it — things we cannot picture yet.
