The agent-domain space filled up faster than I expected
A year ago, "an AI agent registers its own domain" was a novelty. Now it is a category.
AgentDomain lets an agent create an account, fund a wallet, and register a domain with no human in the loop. Sherlock Domains, from Fewsats, is a registrar built for agents with proper SDKs and no browser auth. domainagent.dev does one-turn registration over MCP or HTTP. NameSilo shipped an official MCP server with 80+ methods. And Cloudflare, working with Stripe, now lets an agent spin up an account, pay, register a domain, and get a deploy token in a single flow.
I am building in this space too, so I could pretend I am the first. I am not, and you would find out in about thirty seconds. So instead, here is the more useful observation I keep coming back to.
Almost everyone solved the same problem
Every one of those tools answers the same question: how does an agent complete a registration without a human clicking through a checkout? That was worth solving. It is also, at this point, mostly solved.
The question almost nobody is asking: what kind of domain does a machine actually want? We took the human product, wrapped an API and an MCP server around it, and shipped it to agents unchanged. Same names, same pricing, same assumptions.
But an agent is not a human buyer, and the thing it wants is not the thing a human wants.
Human domains and machine domains are different products
When a person buys a domain, they are paying for memorability. Short. Brandable. Easy to say out loud, easy to type, easy to remember. That scarcity is exactly why a clean four-letter .com costs a fortune. The whole pricing model of the industry is built on human memory being the constraint.
A machine has no such constraint. An agent does not need to remember its domain, spell it at a dinner party, or fit it on a business card. It needs a unique, unambiguous, machine-readable identifier. It may want a lot of them, one per project, per deployment, per sub-agent. For a machine, a long, specific, self-descriptive string is not a downside. It is often better, because it carries meaning and collides with nothing.
So here is the mismatch: the exact domains that are worthless to humans, the long ones, are ideal for machines. And because the industry prices for human memorability, those long domains are treated as leftovers rather than as the right product for a new kind of buyer.
That gap is what I am building AIvikings around.
What that means concretely
AIvikings is a domain registrar that prices for machines, not humans. Long domains, 15 or more characters, the kind an agent is happy to use and a person never would, priced far below what you would pay for them anywhere else.
The agent-native plumbing is there too: a REST API, a native MCP server so any MCP-compatible agent including Claude-based ones can register domains as tool calls, and a ChatGPT connector. But I want to be honest that this part is table stakes now. It is the price of entry to the category, not the reason to choose us. The pricing is the reason.
What honestly works, and what does not yet
Because this is build-in-public, the real status.
Works today:
- REST API for the core domain operations, with API-key auth.
- Native MCP server exposing those operations as tools.
- ChatGPT OAuth connector, working end to end.
- An autonomous demo agent that takes a plain-English goal and drives the full registration loop with no human input.
Not done, and I will not pretend otherwise:
- Billing. Not built yet. The MVP does not charge anyone.
- The final live-registration hardening. Registrations run through a reseller backend I am still moving from validated test mode into a fully proven live path.
- The security and safety work a registrar has to get right before real money and real domains are involved.
That is why there is a waitlist rather than a "register now" button. Some competitors are ahead of me on billing, and I would rather ship this correctly and slightly late than ship something that breaks in front of the exact developers I am trying to win.
Why I am posting this now
I want to argue the machine-domain thesis in public and have you tell me where it is wrong. The whole bet rests on one claim: that long, cheap, machine-readable domains are a real product for agents and not a solution in search of a problem.
So the things I would genuinely like feedback on:
- If you build agents, would you actually want cheap long domains, or is short-and-memorable still what you reach for out of habit?
- How many domains does a real agent workload need? One, or many?
- What would make you trust a new registrar enough to route a real registration through it?
I will keep posting as this moves from pre-launch to live: the live-registration proof, the real pricing numbers, the launch, and the parts that go wrong.
If the machine-domain idea resonates, the waitlist is at https://aivikings.ai. I read every reply here too.
Top comments (0)