On July 28, Anthropic announced the biggest revision of the Model Context Protocol since streamable HTTP: MCP spec version 2026-07-28. The headline change is that MCP moves from a bidirectional, stateful protocol to a request/response model — and around that core, two formally versioned extensions (MCP Apps and MCP Tasks) plus a hardened OAuth story. Here is what actually changed, and what it means depending on where you sit.

What changed, in plain language

  • Stateless core. Servers no longer manage long-lived sessions. Every interaction is a request and a response, which makes MCP servers deployable on serverless and edge infrastructure without session juggling.
  • MCP Tasks. A standard way to run long-running work: start a task, check its status, fetch the result. Until now, everyone hand-rolled their own pattern for anything that took longer than a request timeout.
  • MCP Apps. Servers can render interactive UI directly inside the conversation — not just text results, but actual interface.
  • Authorization hardening. Auth now aligns with production OAuth 2.0 and OIDC deployments, including enterprise identity systems like Okta and Entra.

The scale behind the announcement is worth noting: Anthropic reports over 400 million monthly SDK downloads — a 4x increase this year — and more than 950 servers in Claude’s connectors directory. MCP is not an experiment anymore; it is infrastructure.

If you build an MCP server

This update is almost entirely good news. The stateless core removes the hardest part of hosting a remote server — session management — and makes a plain serverless function a first-class deployment target. If you have been putting off building a server because of infrastructure complexity, that excuse just got thinner.

The practical to-do list: check which spec version your SDK targets, plan to support the new version while clients migrate, and if your server does anything slow (crawling, large exports, batch checks), look at MCP Tasks instead of your own polling workaround. And if your OAuth implementation deviates from standard flows — custom redirect-host allowlists are the common offender — the new spec is your reason to clean that up. In our own probing of official remote servers for the gate directory, non-standard authorization details were the single biggest reason a server failed to connect from a standards- compliant client. Spec alignment fixes that class of failure for everyone.

If you use MCP servers

Rollout is gradual (“across Claude products soon”), and the ecosystem will spend months in a dual-version world: some servers on the new spec, most still on the old one. Well-behaved clients and gateways will speak both, so day to day you should notice little — until the new extensions arrive in your client. MCP Apps in particular will change how tools feel: results that render as interface instead of text walls.

One thing the update does not change: a server that can render UI in your conversation and run long tasks in the background is a server you want vetted before it gets there. The trust questions — what is this server allowed to do, and what did it actually do? — get more important as servers get more capable, not less. A richer protocol raises the ceiling for useful servers and for misbehaving ones alike; that is exactly the gap scanning and per-tool policies exist to cover.

If you run tool access through a gateway

For gateways, statelessness is a simplification: proxying request/response is cleaner than brokering long-lived sessions in both directions. The dual-version window is where a gateway earns its keep — it can speak the new spec to clients while translating for upstream servers that have not migrated yet, so users never have to care which version a given server runs.

The two extensions matter here too. MCP Tasks standardizes exactly the pattern that long-running, outcome-oriented servers need — the kind of specialized agents you hire for a result rather than tools you call for a lookup. And MCP Apps opens an interesting door for governance: approval prompts that render inside the conversation instead of in a separate dashboard. “This tool wants to send an email — allow once?” as native UI, at the moment it matters.

The short version

  • Servers get radically easier to host — expect more of them, faster.
  • Long-running work and in-conversation UI are now standard vocabulary, not hacks.
  • OAuth alignment should retire a whole class of connection failures over time.
  • More capable servers make vetting, policy, and logging more important, not less.

Net: the protocol grew up. The tooling around it — discovery, trust, governance — is where the next year happens.