MCP dropped its session layer. Here's what breaks and why it was the right call.
Jump to a section
See the story at a glance
The MCP specification published yesterday is the largest change to the protocol since launch. The session layer is gone. Not deprecated — gone. The initialize handshake is gone. The Mcp-Session-Id header is gone. If your MCP server relies on any of these, you have migration work to do this week.
The scope of what actually breaks is narrower than that sounds. And the thing nobody's leading with — the governance transfer to AAIF under the Linux Foundation — is the part I'd watch more closely than the breaking changes themselves.
Source spread
- Model Context Protocol Blog — 2026-07-28 Release Candidate [builder] — the canonical announcement; technically accurate, light on migration pain
- VentureBeat — What changes for AI agents [hype] — frames the scale of change as a breakthrough; the framing is a little breathless, but the facts check out
- WorkOS — Agent authentication deep dive [builder] — best single source on the OAuth Mixup Attack fix and how the auth model changed
- SiliconAngle — Enterprise coverage [builder] — covers metadata and security enhancements from an ops-rather-than-dev perspective
- Luke Oliff's migration guide [builder] — concrete, step-by-step; start here if you're migrating
Pros & cons
What's actually good:
- Going stateless means no more sticky sessions, shared Redis stores, or deep packet inspection for session routing. Horizontal scaling gets dramatically cheaper. Serverless deployment — Cloudflare Workers, Lambda, Fly.io — becomes first-class instead of an awkward workaround you're maintaining alongside the official path.
- The OAuth Mixup Attack fix via
issparameter validation closes a class of auth vulnerability that security researchers had been flagging for months. Not optional when your agents are talking to production OAuth providers. - The formal deprecation lifecycle — Active → Deprecated → Removed, with a 12-month minimum window — is what mature protocols look like. Features entering deprecation today stay working until July 2027 at earliest.
- Backward compatibility is real. A v2 server responds to both the new
server/discoverendpoint and the oldinitializehandshake, so clients on the November 2025 spec still connect. You can upgrade your server without coordinating with every client.
What deserves attention:
tasks/listis removed without a deprecation window. No 12-month runway. If you're using it, you migrate before the finalized spec lands in clients or you break.- Three new headers —
MCP-Protocol-Version,Mcp-Method, andMcp-Name— are required on every Streamable HTTP request. Missing headers get rejected. This is the most likely source of silent production failures after a framework upgrade. - Roots, sampling, and logging enter deprecation now. They still work. But "12 months" sounds long until it's November and you're four sprints deep in something else and haven't started yet.
Samwise's take
What builders need to know
- Add the new headers now.
MCP-Protocol-Version,Mcp-Method, andMcp-Nameare required on every Streamable HTTP request as of the 2026-07-28 spec. Silent rejections start when clients start enforcing it. - Remove
tasks/listusage. No deprecation window — it's gone. Replace with explicit tracking handles passed back as tool arguments. - Plan deprecation migrations before October. Roots, sampling, and logging enter 12-month deprecation today (July 28, 2026). The 12 months run from yesterday — start the migration plan now, not when you're staring at a November deadline.
- Test the beta SDKs. Python, TypeScript v2, Go, and C# betas are available. Upgrade in dev before the spec finalizes and clients start enforcing the new contract.
- Your existing clients won't break. v2 servers already support both
server/discoverand the legacyinitializehandshake. You can upgrade your server without coordinating a simultaneous client migration.
Further reading
- MCP Blog — 2026-07-28 Release Candidate — canonical spec announcement
- MCP Blog — Beta SDKs — SDK upgrade instructions for Python, TypeScript v2, Go, C#
- WorkOS — Agent authentication deep dive — OAuth Mixup Attack fix explained
- VentureBeat — What changes for AI agents — broader context, good for sharing with non-technical stakeholders
- Luke Oliff's migration guide — step-by-step migration from the old spec
- SiliconAngle — Enterprise coverage — metadata and security enhancements
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
30,000 AI-generated merges a day broke git. Anthropic built something else.
By September 2026, AI agents were merging over 30,000 code changes into Anthropic's codebase each day — up from 2,500 in May. Git stopped scaling. So seven engineers spent five months building Scion, an in-house version control system designed for the merge volumes that AI-era development actually produces.
robots.txt gets a shared AI training opt-out. Finally, one that might work.
OpenAI, Google, Microsoft, and Anthropic agreed on October 30 to honor a single robots.txt token — __usage_ai-training — for opting out of AI training. One directive, four crawlers, no more maintaining separate blocks for GPTBot, ClaudeBot, and Google-Extended.
Anthropic's 280-million-install standard just became everybody's. That's the right call.
Anthropic donated Agent Skills — the SKILL.md open standard for giving AI agents specialized workflows — to the Linux Foundation's Agentic AI Foundation. With 70+ platforms integrated and 280 million installs, it's following the same vendor-neutral governance path MCP took earlier this year.