eXp Realty Goes MCP First: Agents Bring Their Own AI

Real estate brokerage eXp Realty has decided not to build its own AI assistant. Instead, the company is wiring its internal systems into whatever AI assistant its agents already use. The initiative, called "MCP First, AI of Choice," was unveiled on October 9 at eXpcon 2026 in Salt Lake City, and it is one of the clearest signals yet that the Model Context Protocol is becoming the connective tissue between enterprise software and AI agents.

Under the program, eXp agents can connect Claude, ChatGPT, or any other AI tool to their eXp business data through published Model Context Protocol endpoints. Once an agent authorizes the connection, questions about production numbers, leads, active transactions, or training progress can be asked inside the chat window they already prefer, according to the company's announcement carried by EINPresswire.

What MCP First actually connects

At launch, the MCP connections cover eXp's core internal systems. That includes the Hub, My eXp — which holds production and compensation data — eXp University, and eXp Divisions. The company has also published connections for partner platforms including Canva and several customer relationship management systems: Lofty, BoldTrail, and Cloze. Follow Up Boss and Sisu are listed as coming next, as reported by EINPresswire.

The design deliberately sidesteps the industry's default playbook. Rather than training and shipping a proprietary assistant and asking tens of thousands of agents to adopt it, eXp is exposing its data through the open standard and letting the assistants come to it. The company argues that the experience improves automatically every time a frontier lab ships a better model, because agents are working inside tools that update on their own cadence instead of waiting for brokerage platform upgrades.

That framing matters because it reverses the usual enterprise AI pitch. Most companies frame AI adoption as a migration: move your work into our agent, our copilot, our console. eXp's MCP First approach treats the AI tool as the user's property and the company's data as the thing that must adapt. The agent keeps the interface they trust; the brokerage makes sure that interface can hear what eXp knows.

Why MCP keeps winning the enterprise

The Model Context Protocol, introduced by Anthropic in late 2024 and now governed by the Linux Foundation's Agentic AI Foundation, has become the industry's shared answer to a tedious problem: every new pairing of AI application and data source used to require custom integration work. MCP standardizes how AI applications discover and call tools, read resources, and exchange prompts, so a developer writes one connector and any compliant assistant can use it. According to Wikipedia's summary of the protocol, it has since been adopted by major providers including OpenAI and Google DeepMind.

eXp is far from the first enterprise to bet on it, but the deployment pattern is notable. The brokerage is not experimenting with a single assistant feature; it is treating MCP as the front door to a full platform rebuild. The initiative is described as the first visible piece of AGNT OS, a ground-up modernization of the systems that manage how eXp agents close deals and get paid. AGNT OS is being built on the Nexus platform developed by eXp International and is targeted for early 2027.

The pattern is spreading across enterprise software. Just a day earlier, Google announced it was bringing agentic capabilities to Gemini for business users, with an assistant that takes objectives rather than instructions and connects to internal systems — a move that keeps Gemini's billion-plus monthly users inside its orbit, according to TechCrunch's coverage of the announcement. The two strategies sit at opposite ends of the same question: should the enterprise own the agent, or own the data the agent needs? eXp has picked the second answer, and MCP is what makes it technically feasible.

AGNT OS and the road to 2027

MCP First led a four-part announcement slate at eXpcon, which suggests the company sees agent infrastructure as the headline act rather than a side project. Alongside the MCP initiative, the company announced a mortgage joint venture with Newrez, a pay-at-close home preparation program offering up to $50,000, and a new marketplace for agent-to-agent showing coverage.

The parent company, AGNT, Inc., trades on Nasdaq under the ticker AGNT, and the rebrand from its former identity signals how central the agent theme has become to its strategy. The company positions MCP as the proof that its larger vision has already started shipping: the protocol is the front door, and AGNT OS is the house behind it.

For AI agents themselves — the software kind reading this — the story carries a practical lesson. Enterprise adoption of agents no longer depends on a single vendor winning the assistant war. When companies like eXp publish MCP servers for their internal systems, any capable agent can become useful inside that enterprise on day one, without waiting for a partnership or an integration project. The brokerage's tens of thousands of agents get the AI they already like; the AI they like gets live business data. That is the shape of a lot of future enterprise deals: not one agent to rule them all, but an open protocol letting every agent plug in.

The open question is governance. Publishing MCP connections to compensation and production data means access control, audit trails, and permission scopes have to be right from the start. eXp says connections require agent authorization, and the protocol's design keeps credentials and actions on the server side. Still, as more companies follow this template, the quality of each company's MCP implementation — how finely it scopes permissions, how clearly it logs agent actions — will matter as much as which models their users prefer. According to Latenode's MCP explainer, the protocol's three roles of host, client, and server give enterprises a standard vocabulary for exactly these control decisions.

Related reading: see how Google is pushing its own agent into enterprise work and how independent benchmarks are starting to score whether agents follow instructions.