Bring Your Own AI Is the New Front Door

For two years nearly every software company did the same thing. They shipped a chatbot with a sparkly icon in the corner and a cute name, and called it their AI strategy. This week one of the biggest real estate brokerages in the US looked at that playbook and passed.

At its eXpcon event on 8 October, eXp Realty said it is not building another branded assistant. Instead, its agents can give the AI they already use (Claude, ChatGPT or Gemini) access to their eXp data through the Model Context Protocol, or MCP. Production numbers, leads, transactions, training progress and CRM records all appear in the chat window the agent already works in.

eXp calls it "MCP First, AI of Choice." To me it looks like the moment the chatbot stopped being the product.

What eXp actually shipped

Take away the branding and there are three parts:

• A connection, not an app. Agents let their own assistant read their brokerage data over MCP. There is no new login screen and no new interface to learn.
• A prompt library. Once connected, agents get ready-made prompts such as "my production over the past year" or "CRM contacts with birthdays this month."
• A bigger roadmap. MCP is described as the first piece of a wider rebuild called AGNT OS, due in early 2027.

eXp's CEO set the move against competitors' "walled proprietary gardens." To be fair, the company could not share adoption numbers, so this is a statement of strategy, not proof that it worked. The strategy is what interests me.

Stop building hair dryers. Install outlets.

Think about a good hotel. It doesn't make you use its own hair dryer, its own phone charger and its own laptop. It gives you a power outlet and lets you bring your own devices. The outlet is standard, so whatever you bring just works.

MCP is that outlet for AI. For years software companies built hair dryers: their own assistants, trained on their own data and stuck inside their own apps. eXp stopped doing that and put in outlets. Whichever assistant the agent prefers, it plugs in and reaches the same data.

The customer's assistant is becoming the front door. Your product becomes the rooms behind it.

That changes what "having AI" means for a software product. It is less about the chat box on your screen and more about whether someone's assistant, wherever it runs, can reach your data and actions and get a correct answer.

This is bigger than one brokerage

If eXp were alone, you could call it a clever marketing angle. It isn't alone. The same week, Atlassian announced a rebuilt MCP offering that covers more of its products, and it says the layer handles more than 15 million MCP calls a day. That is people at work sending real questions from their assistants into the tools where the work actually lives, every day, at scale. It is not a pilot.

Oracle is coming at the same idea from the enterprise side. Its new Fusion Claw runtime lets a frontier model (Gemini or OpenAI can start a workflow) do the planning, then hands execution to deterministic enterprise computation for volume and repeatability. Each run is a "Claw Outcome" with run-level permissions, approvals and audit controls. It launched with 25 new agentic apps, bringing Oracle's Fusion agentic portfolio to 75.

The layer is different but the lesson is the same. The AI model is a guest you invite in under clear house rules. It isn't the house.

Why "AI of choice" is the smarter bet

I think this pattern will spread, for three reasons:

1. People have already picked an assistant. Asking someone to learn a second, weaker one inside your app adds friction. Meeting them in the assistant they already use takes it away.
2. Models change every few months. A brokerage, a project-management company or an ERP vendor will not out-build the frontier labs on the model. They don't need to try.
3. The data and the actions last. Whoever holds the transactions, the tickets or the ledger stays essential however the chat window changes.

The sleeper feature: a prompt library

The part of eXp's announcement I like most is the least flashy one. Anyone who has stared at an empty chat box knows the problem: you have a powerful tool and no idea what to ask it.

A prompt library fixes that. It works like a menu. You don't need to know what the kitchen can cook if someone hands you a list of good dishes. "Show me my production over the past year" is a much better first experience than a blinking cursor. In this model, the prompt library is the onboarding.

What to watch next

Some practical takeaways, depending on where you sit:

• If you use software at work, expect more "connect your AI" buttons over the coming months. Before you click, check what you are granting. Can your assistant only read data, or can it take actions? Oracle's focus on approvals and audit trails is a good guide to what serious access controls look like.
• If you build software, ask a simple question: could a customer's own assistant answer a real question about their account using your data today? If not, that gap matters more every month.
• If you just follow the industry, the AGNTCon + MCPCon North America livestream on 22 October is a good place to see where the standard is heading.

The chatbot era taught every company to build its own voice. The next era rewards companies that make their data easy for any voice to reach. eXp, Atlassian and Oracle each made that bet in public this fortnight, and it looks like a good one to me.