MCP is Sticking

“Protocols are dead.” Sure—and yet every cloud giant quietly aligned on MCP. Funny how “too slow” turns into “table stakes” once the pain is real.

As of July 2025, MCP isn’t hype. It’s boring infrastructure. That’s the tell.

The Room Where It Happened

On Feb 22, Mahesh Murag (Anthropic) spoke about MCP to ~200 AI engineers. It became one of the top videos on the AI Engineer YouTube.

The kicker: OpenAI had a talk at the exact same time… and nobody showed.

I don't know if the conference organizers meant to schedule both at the same time, in adjacent rooms, but it was an incredible A/B test. Do you go to the Anthropic talk (A) or the OpenAI one (B)? Can't pick both.

Almost everyone picked Anthro.

The hallway was buzzing afterward. Developers were already sketching out integrations on napkins. The energy was palpable—that rare moment when you see something that just clicks architecturally.

And Then OpenAI Got the Memo

Timeline of events that definitely aren't coincidental:

• Feb 22: Anthropic packs the house at AI Engineer NYC
• March: OpenAI announces MCP support (quietly, but officially)
• April: Rumors start flying about OpenAI acquiring Windsurf
• May: They release Codex & Codex CLI with MCP integration
• July: Microsoft goes full MCP-pilled (Kevin Scott's been tweeting about it non-stop)

What did OpenAI see? Developer gravity around Claude. You don’t fight that head‑on; you redirect it. Hence: support the protocol developers are adopting.

If I were in Sam's shoes, I'd definitely be thinking:

"We know our models and products are excellent, but if a critical mass of developers are using Anthropic's stuff, we'll be typecast into being just the chat/search/intelligence giant for consumers. We can't let that happen. We have a $500 billion Stargate cluster that we need to feed tokens. Nobody uses more tokens than devs. We need to (re)-capture this market."

Why This One’s Different

Most protocols die because they're designed by committee, over-engineered, and solve yesterday's problems. MCP is different:

• Timing: It arrived exactly when we needed it—when every team was building their own terrible agent frameworks
• Simplicity: The spec is readable by humans (shocking, I know)
• Real adoption: Companies are actually using it, not just announcing support
• Competitive pressure: When OpenAI, Microsoft, and Anthropic all support something, it's real

The pain: giving agents tool access without bespoke glue. Pre‑MCP, every framework rolled its own, and everyone re‑learned the same lessons the hard way.

What This Means for You

Being conversant in MCP is already valuable. Learn‑it‑this‑weekend valuable.

Your action plan:

• Start building: Pick a tool you use daily and write an MCP connector for it
• Read the spec: It's surprisingly approachable—actual humans can understand it
• Experiment with combinations: The real power is in chaining multiple tools together

The companies that get ahead of this curve will have a significant advantage. The developers who become MCP-fluent will be in high demand.

Shipping Check

This isn’t just another protocol war. When competitors converge on a boring standard, market forces beat ego. Rare—and durable.

The Feb 22nd event was the canary in the coal mine. Developers voted with their feet, and the big players noticed.

Time to (keep) building. The MCP train is leaving the station, and you want to be on it.

Next step: Build one MCP connector this weekend. Smallest useful thing. Document what surprised you. Share it.

The future belongs to the builders who understand how to make AI agents actually useful. MCP is how we get there.

If you have insights or feedback on MCP adoption, feel free to share your thoughts in the comments.