The Model Context Protocol just shipped its biggest rewrite since launch. On July 28, 2026, MCP’s maintainers finalized the 2026-07-28 specification — the standard that lets AI models securely connect to external tools, files, and services — and the headline change is that the protocol is now stateless at its core, a shift aimed squarely at running AI agents at real production scale.
Quick facts
- The MCP 2026-07-28 specification finalized on July 28, 2026, described by its maintainers as the largest revision since the protocol launched.
- The core protocol is now stateless, removing the requirement to track a session ID across requests so any server instance can answer any request.
- Tasks (long-running operations) and MCP Apps (server-rendered UIs) move out of the core spec into a new formal Extensions framework.
- Roots, Sampling, and Logging are formally deprecated, along with Dynamic Client Registration; deprecated features stay functional for at least 12 months.
- Adoption is enormous: MCP’s Tier 1 SDKs see nearly half a billion downloads a month, with both the TypeScript and Python SDKs individually crossing 1 billion total downloads.
What MCP actually is, for anyone catching up
MCP is the plumbing that lets an AI model reach into your calendar, your database, or an internal company tool without an engineering team building a custom connection for every single service. Instead of every AI product inventing its own integration format, an MCP server exposes a standard set of capabilities — callable tools, readable resources, reusable prompt templates — that any MCP-compatible client can use the same way. It’s become the de facto standard for wiring AI agents up to real-world systems since Anthropic introduced it in late 2024.
Why “stateless” is the change that matters most
Under the previous specification, an MCP client and server tracked a session together using a session ID header, meaning the same server instance generally needed to handle every request in a conversation. That’s a real constraint at scale: it makes load balancing harder and forces more state to live in one place. The new stateless core removes that requirement entirely, achieved through six separate Specification Enhancement Proposals working together, according to the official MCP blog. In practice, that means any request can now be answered by any available server instance behind ordinary HTTP load-balancing infrastructure, which is exactly the kind of infrastructure most companies already run everything else on.
Anthropic’s David Soria Parra, one of MCP’s lead maintainers, called it the most substantial change to the specification, according to reporting from The Register, probably since authorization was added. He was direct that this isn’t a drop-in upgrade: the underlying data transfer mechanism has been rebuilt, and “a lot of things that made MCP are gone” in their old form.
What’s deprecated, and what that means for you
Roots, Sampling, and Logging are formally deprecated in the new spec, along with Dynamic Client Registration, which is being replaced by a newer Client ID Metadata Document approach. None of this breaks overnight — MCP’s formal deprecation policy guarantees deprecated features keep working for at least 12 months — but it’s a real migration project for anyone maintaining MCP servers or clients, not a background update you can ignore. Worth noting explicitly: servers built on the new revision aren’t guaranteed to work with older clients, and vice versa, so mixed-version environments need real compatibility testing rather than assumptions.
The enterprise piece: centralized authorization
Separately, on July 6, 2026, MCP’s Enterprise-Managed Authorization extension reached stable status, according to InfoQ. It lets organizations control access to MCP servers centrally through their existing identity provider, replacing per-server consent prompts with a sign-in-once flow. For any company managing dozens or hundreds of internal MCP servers, that’s the difference between a security team that can actually audit access and one drowning in individual approval requests.
Why the download numbers matter
Close to half a billion downloads a month across the official SDKs, with the TypeScript and Python SDKs each individually past a billion downloads total, is the real signal here: MCP isn’t a promising standard anymore, it’s already the substrate a huge share of production agentic workflows run on. That’s exactly why a breaking change to the core transport layer is a genuinely big deal rather than routine protocol housekeeping — it touches an enormous, already-deployed base of agent frameworks and integrations built on the old assumptions.
Common questions
Do I need to upgrade immediately? No. Deprecated features remain functional for at least 12 months under MCP’s formal deprecation policy, giving server and client maintainers real runway to migrate rather than a hard cutover.
Will my existing MCP server keep working with newer clients? Not guaranteed. Because the transport layer itself changed, cross-version compatibility needs explicit testing rather than assumption — treat it the same way you’d treat any breaking API version change.
What replaced Tasks and MCP Apps in the core spec? Nothing replaced them — they moved out of the core protocol into the new formal Extensions framework, meaning they’re still fully supported, just structured as optional add-ons rather than baked into the base spec everyone must implement.
Key takeaway
If you build or host MCP servers, this isn’t optional reading: audit your implementation against the 2026-07-28 changelog, check whether you depend on Roots, Sampling, Logging, or Dynamic Client Registration, and test compatibility explicitly rather than assuming your existing clients and servers will keep talking to each other across the version boundary.
