You’re in a meeting. Someone from engineering says, “We’re not using MCP.” A few heads nod. You nod too, because everyone else is nodding and you’d rather not be the person who asks what MCP stands for in front of eleven people.
Later, you look it up. Model Context Protocol. A standard way for AI agents to connect to tools and data. Sounds useful? So why did your team say no?
If this is you, I have good news. The “no MCP” conversation happening across tech right now is not actually a story about something failing. It’s a story about something growing up in public, which is messier and more interesting.
What Happened in March
On March 11, 2026, at an event called Ask 2026, Perplexity’s CTO Denis Yarats announced that the company was moving away from MCP. His reason was specific and technical: MCP tool definitions were eating 72% of the context window.
Let me translate that, because it’s the whole ballgame.
An AI model has a limited amount of working memory for any single conversation. Think of it as a desk. Everything the model needs to consider — your question, the relevant documents, the instructions, the list of tools it’s allowed to use — has to fit on that desk.
Yarats was saying that the instruction manuals for the tools were covering 72% of the desk. The agent spent most of its memory just remembering what it was capable of, leaving a sliver of space for the actual work.
That’s a real problem. It’s also a very different problem from “this technology is bad.”
Why Everyone Jumped to the Funeral
Within days of that announcement, the “MCP is dead” takes started. This is a pattern worth recognizing, because you’ll see it again with the next protocol, and the one after that.
One well-known company with specific needs hits a specific limitation. That gets reported. The reporting gets compressed into a headline. The headline gets compressed into a vibe. And then the vibe shows up in your meeting as “we’re not using MCP,” delivered with the confidence of settled fact.
Perplexity’s situation is not every situation. They run a search product at enormous scale where every token of context has a cost. A company with a different shape of problem might never hit that wall.
The Part the Headlines Missed
On July 28, 2026, MCP released a new specification. The headline change is that MCP is now stateless.
Here’s what that means without the jargon. The old version worked like a coat check. Your agent would start a session, get a ticket, and then every follow-up request referred back to that ticket. The server had to remember who you were and what had already happened.
The new version removes the coat check entirely. There’s no handshake and no ticket. Every single request carries everything the server needs: protocol version, client identity, capabilities, all bundled in.
Each request is self-contained. The server doesn’t have to remember anything.
If you’ve ever wondered why engineers care so much about this kind of thing, it’s because stateless systems are dramatically easier to run. You can add more servers without worrying about which one holds your session. A server can crash and restart without dropping your work. Deployment over plain HTTP gets simpler.
The spec update also added extensions for interactive apps and long-running tasks. Cloudflare’s write-up on the new generation of MCP noted the release came alongside updated SDKs for TypeScript, Python, Go, and C#. The official MCP blog framed the stateless shift as a significant step for enterprise scalability, specifically removing deployment friction.
Why Enterprises Didn’t Walk Away
Large organizations kept paying attention to MCP through all of this, and the reason has little to do with speed or elegance. It comes down to three unglamorous things: authentication, audit trails, and governance.
- Authentication means proving the agent is who it claims to be before it touches anything.
- Audit trails mean a record exists of what the agent did, when, and with what data.
- Governance means rules about what’s allowed, enforced consistently rather than hoped for.
A clever custom integration can be faster than a standard. It usually can’t answer a compliance officer’s questions. That’s why MCP’s value proposition in these areas kept it on the table for enterprises even while the “dead” takes were circulating.
What to Do With This
If your team said no to MCP, that might be exactly right for your situation. Context window costs are real.
But if the no came from a headline rather than an evaluation, the ground has shifted since March. The thing being rejected in the spring is not the thing that exists now.
The useful question isn’t whether MCP is alive or dead. It’s narrower: what are we actually connecting, how much context does that cost us, and who needs to be able to audit it later? Those answers are yours, not Perplexity’s.
🕒 Published: