Connecting Claude to your CRM with MCP

Every real estate platform has an API. That’s been true for a decade and it has changed almost nothing for the average agent, because an API is a promise that somebody could build an integration — if they had a reason, a budget, and a developer.

MCP is different in a way that took me a while to appreciate.

The actual distinction

With a REST API, someone writes code that says: when this happens, call this endpoint, map these fields, handle these errors. That code is a product decision. It exists because a company decided the integration was worth building, which means it exists for the popular combinations and not for yours.

With MCP, the server publishes a list of tools and what each one does, in plain language. The AI client reads that list and works out which one it needs for what you asked. There’s no integration to write, because the “integration” is the model reading a menu and choosing.

That sounds like a small distinction. In practice it means the long tail suddenly works. Nobody at our company had to anticipate that you’d want to ask which contacts from a specific source haven’t been touched in six weeks. The model figures out it needs contact search plus timeline data and goes and gets it.

What it feels like

The honest version: about 80% of the value is not switching tabs.

You’re drafting something. A listing description, an email to a client, a plan for the week. You need to know what’s actually on your calendar tomorrow, or whether that seller ever replied. Normally that’s: leave what you’re doing, open the CRM, find the record, read it, come back, remember what you were saying.

With the connector wired up, you ask, and the answer arrives in the conversation you were already having.

The other 20% is questions you’d never bother opening a CRM to answer. “Who’s gone quiet this month that I should probably call?” is a genuinely useful question that almost nobody asks their CRM, because framing it as a filtered smart list is more work than the answer is worth. Asking it in a sentence is not.

Setting it up

If you’re on Claude Code it’s one command:

claude mcp add --transport http brokeragent \
  https://mcp.brokerandagent.com/mcp \
  --header "Authorization: Bearer bak_your_workspace_key"

Claude Desktop takes the same URL and header as a custom connector. Any MCP client that speaks streamable HTTP works the same way — the protocol is open and we didn’t invent any of it.

The key comes from Settings → Integrations in your workspace. When our OAuth server ships this becomes “sign in and approve” with no key to copy, which is better in every way, but the key works fine today.

What we let it do, and what we don’t

This is the part I’d actually read carefully if I were evaluating any AI connector, ours included.

It can read fairly generously: workspace info, contact search, individual contacts, contact timelines, pipeline summary, tasks, appointments, transactions, automations, today’s schedule. Reading is low-risk and it’s where most of the value is.

It can write, narrowly: notes, tasks, leads, appointments, and enrolling someone in an automation. Light capture. The things you’d want to jot down without opening anything.

It deliberately cannot send texts or emails to your clients, delete anything, move deals between stages, or touch settings, billing or your team. Those are absent on purpose, not on a roadmap. An AI working on your behalf shouldn’t be able to do the things you’d want to look at first, and “send a message to a client” is squarely in that category regardless of how good the model is.

Everything runs as you. Same API your own session uses, same key, same permissions, same plan limits. There’s no privileged back door, which matters because a connector with elevated access is a connector that can do more damage than you can.

Where this is going

I’ll be careful about predictions, but one seems safe: the assistant you already use becoming the front door to your business software is going to be normal within a couple of years, and the products that get there early will feel obvious in hindsight.

What I’m less sure about is how much of the work moves there versus how much stays in the app. My guess is that reading and asking move almost entirely, and the actual doing — building the CMA, working the pipeline, running the transaction — stays where it is, because those benefit from a real interface. Which is roughly how we’ve built it: MCP is the open door, and the in-app assistant does the heavy lifting.

As far as we can tell we’re the only real estate platform shipping a native MCP server as of September 2026. That’ll change, and it should. It’s an open standard and the whole point is that everyone can implement it.

If you build things and you want to poke at it, the developer page has the full tool list and the REST API and webhook details. It works on the free plan, because gating the developer surface behind a paid tier is how you end up with no developers.