Zap Imóveis, OLX, Viva Real: visual navigation platforms — filters, forms, pages to click through. The architecture assumes a buyer who knows where to browse.

Language models don't browse. They query. When an AI agent receives the question — "find a 2-bedroom apartment in Moema for R$800K" — it doesn't open Zap. It consults sources it can read: structured, semantic data accessible via protocol.

Brazilian brokerages aren't on that list.

Model Context Protocol (MCP) is the standard that defines how AI agents integrate external data sources. In less than 12 months, it was adopted by Anthropic, OpenAI, Google DeepMind, and Microsoft. In December 2025, it was donated to the Linux Foundation — becoming open infrastructure maintained by the same consortium that oversees HTTP. This isn't a bet on any one company. It's protocol infrastructure.

Today: 97 million monthly SDK downloads. More than 10,000 servers in active production.

The correct analogy isn't with apps. It's with SEO between 1998 and 2003.

In 2002, those who understood that websites would need to be readable by search engines had a two-year head start on the market. Semrush, Moz, Ahrefs — all born from that window. Those who built position in that moment defined the next decade of digital distribution.

This is the same curve. With one difference: the clock runs faster.

That's the right question. Here's the complete answer.

Zap Imóveis' revenue model depends on a single premise: that the brokerage has no alternative distribution channel. R$4,200/month per brokerage — multiplied across tens of thousands of clients — is the number that sustains the entire business. A product that creates direct distribution, without portals as intermediaries, destroys that premise.

Portals cannot build this from within. The structural incentive prohibits it.

There's also a technical argument, and the American market has already proven it: Zillow, Redfin, Realtor.com, and Homes.com all launched AI in 6 months — and all built within their own walled gardens. Closed features, closed ecosystems. None opened inventory to the LLM ecosystem. That's not oversight. That's business model.

MCP is an open standard. If Zap builds an AI feature, it's closed — it works within their platform, accessible only by the models they integrate. Roofhound connects each brokerage to the entire ecosystem: Claude, GPT, Gemini, and any agent that adopts the protocol. That's the difference between a walled garden and open protocol infrastructure.

The data that matters isn't the volume of pilots. It's how they were acquired.

Each pilot came from a personalized demo — a live search showing the brokerage's properties not appearing on ChatGPT. No marketing campaign. No aggressive discounting. No follow-up sequences. The same approach generated the first response in under 48 hours.

That's the signal that distinguishes real traction from forced pipeline: the brokerage wasn't looking for a solution. It didn't know it had the problem. When it saw it, it asked to join.