MCP Is Becoming the USB-C of AI — And There's a TCP/IP for Agent Identity Too

Two things happened this week that, on their own, are interesting infrastructure stories. Together, they're something bigger: the foundations of how the AI agent economy will actually work.

The first: the Model Context Protocol (MCP) is about to ship its biggest update yet on July 28, locking in its position as the standard connector between AI models and the tools they use. The second: Vint Cerf, co-creator of TCP/IP, retired from Google last week and immediately joined Innovation Labs to push DNSid through the IETF, a standard for giving every AI agent a verifiable, domain-anchored identity.

One is about how agents talk. The other is about who's doing the talking. If you're building with no-code tools, or picking platforms for your team, you need to understand both.

MCP won

I remember when MCP launched in November 2024 and it felt like just another Anthropic project. Eighteen months later, it's the standard. OpenAI adopted it in March 2025. Google DeepMind followed in April. In December, Anthropic donated it to a Linux Foundation fund co-founded with OpenAI and Block. The thing has its own foundation now.

The numbers back it up. Over 10,000 public MCP servers exist, with 9,652 in the official registry. SDK downloads are hitting 97 million a month. Stacklok's 2026 survey found 41% of software organisations in production with MCP servers. Salesforce processed 4.5 million MCP calls within weeks of launching their Headless 360 integration.

And the July 28 spec release is serious engineering: stateless protocol core, first-class Extensions, server-rendered UIs through MCP Apps, long-running Tasks, proper OAuth alignment. This isn't a scrappy open-source side project anymore. It's infrastructure.

The USB-C comparison is a bit worn by now but it holds. Before MCP, connecting an AI to your CRM, database, and docs meant custom integrations for every combination. Now? One standard. Write an MCP server once, every MCP-compatible AI can use it. For no-code platforms, this changes the game. If your platform supports MCP, you don't need to build individual integrations for every AI tool. You wire up once and the whole ecosystem of MCP servers becomes available.

Vint Cerf wants to give agents ID cards

Now the identity side. This is where it gets properly interesting.

Cerf left Google after 20 years. His first move? Not a cushy board seat at some AI giant. He went to a DNS registry company's innovation lab to build the identity layer for autonomous agents.

DNSid works on a dead-simple principle: every AI agent gets a Fully Qualified Domain Name anchored to a domain controlled by whoever's accountable for it. The domain owner publishes cryptographic keys, a lifecycle log, and operational status in DNS TXT records. Any system can verify an agent's identity independently, without phoning home to a central authority.

The IETF draft landed April 30. Innovation Labs is already trialling it with unnamed hyperscalers and identity companies. Cerf's argument, as he told TechCrunch, is basically the TCP/IP playbook: "Company X uses agent Y's technology, and company A uses agent C's technology, and then they don't interwork with each other. Nobody can do everything that you might want every agent to do, so we're going to have to rely on the pressure coming from the users."

That's not a prediction. That's someone who watched this exact dynamic play out once before and knows how it ends.

Here's what I think is the most important detail: neither OpenAI nor Anthropic has signed on. DNSid is being built by a DNS registry company that has no AI business and doesn't want the registration data. As Innovation Labs interim CEO Allie Kline put it, "There's a lot of organ rejection to a hyperscaler releasing a standard and having that proprietary data."

Open infrastructure, no vendor lock-in, pushed through the IETF. Cerf is playing the long game, same as he did in the 70s.

Two halves of the same problem

MCP and DNSid solve complementary problems. MCP standardises what an agent can do and how it asks for things. DNSid standardises who the agent is and who's accountable when something goes wrong.

Think about what happens when an enterprise deploys AI agents in production. An agent from your procurement system needs to query a supplier's inventory API. Another agent from finance needs to authorise a payment. A customer's agent needs to negotiate a return with your support agent.

Right now, every one of those interactions is a bespoke integration with custom auth, custom logging, and no shared way to verify who's on the other end. MCP handles the "how do we talk" part. DNSid handles the "who are you, really" part. Together they're the transport layer and the identity layer of the agent internet.

The parallel to TCP/IP isn't perfect. MCP is an application-layer protocol, not a transport protocol. But the architectural role is the same: open, interoperable standards that nobody owns, adopted because they solve a real problem that proprietary alternatives can't.

What this means for no-code builders

If you're building with no-code or low-code tools, or evaluating platforms for your organisation, these two standards should be on your radar. Here's why.

First, the platforms that support MCP get access to an ecosystem that's growing explosively. Make.com already has an MCP gateway connecting AI agents to their scenario builder. The 10,000-plus servers cover everything from databases to CRMs to developer tools. A no-code platform that doesn't support MCP is like a phone that only works with one brand of charger. You'll outgrow it fast.

Second, agent identity is about to become an enterprise requirement. When your finance team's AI agent initiates a £50,000 payment, someone is going to ask: which agent did that, who authorised it, and where's the audit trail? DNSid gives you cryptographic answers to those questions. Platforms that build identity into their agent architecture from day one will have an enterprise trust moat that bolt-on AI features can't replicate.

This is where the distinction between platforms that were built agent-native and those that layered AI on top of existing tools starts to matter. Agent-native architecture means identity, audit trails, and permission models are part of the design, not afterthoughts. It's the difference between a building with a security system and a building where someone taped a camera to the door.

What to ask your vendors

When you're evaluating a no-code or AI platform this year, ask two questions:

• Do you support MCP? Not "do you have AI features". Specifically: can your agents connect to external MCP servers, and can external agents use MCP to call your platform's tools?
• What's your plan for agent identity? If they look confused, that's your answer. If they mention open standards, ask if they're tracking DNSid or the IETF work. You're not looking for a commitment to a specific draft. You're looking for evidence they understand this problem exists.

The platforms that have good answers to both questions are the ones that will be relevant in three years. The ones that don't are betting the agent economy won't happen, or that they can catch up later. I wouldn't take that bet.

Cerf told TechCrunch he doesn't think the agent economy is inevitable. But he does think people will try: "We are fundamentally lazy creatures, and if we find a way to have an agent do something for us, we're very likely to choose to do that."

He's right. And when those agents start transacting with each other at scale, the platforms that speak MCP and carry DNSid credentials will be the ones anyone trusts to join the network.