Can Your CRE Data Platform Talk to Claude? What MCP Means for Tenant-Rep Brokers

MCP stands for Model Context Protocol. It is an open standard that lets an AI assistant such as Claude or ChatGPT call outside tools and read outside data in a structured way, instead of guessing from whatever it memorized during training.

Anthropic released it in late 2024 and, in December 2025, donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. At that point the protocol reported over 97 million monthly SDK downloads and roughly 10,000 active servers, with first-class client support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. Those adoption figures come from the protocol's own maintainers, so treat them as a platform claim rather than an independent audit, but the governance move is a matter of public record.

The mechanics matter less than the shape. A vendor stands up an "MCP server" that exposes a short list of named tools. Your assistant calls those tools with your credentials, gets back real records, and reasons over them in conversation. The vendor keeps its access controls and billing. You keep your assistant.

DealGround is a prospecting platform built on title data for roughly 160 million US properties, plus tenant and lease data extracted by AI from broker offering memoranda, plus an owner-research layer that works from an LLC on title through Secretary of State filings to a named person with a phone number. It is aimed at investment-sales and leasing brokers, and its demos lean retail and industrial.

On June 1, 2026, DealGround published the technical details of its MCP server. Eight tools: property search, a schema-describe helper, owner-portfolio lookup, LLC unmasking with contact and confidence, batch status, saved-list creation, saved-list analysis, and a liveness check. Every call is authenticated as the connected user through Auth0 and respects the same permissions as the web app. Owner resolution spends credits, so the vendor explicitly asks AI clients to confirm before invoking it. Saved lists created by an AI client get an attribution suffix so a team can tell which lists a human built.

In a September 2 webinar covered by the AI for CRE Collective, CEO Dan Mosher demonstrated the connection live: he asked Claude for the five lowest-cap-rate McDonald's properties marketed in California over the previous 18 months and their ownership contacts. Claude handled the conversation. DealGround ran the searches and returned the records. The results were saved back into DealGround as a list.

That is the first time we have seen a CRE data vendor treat "the assistant the broker already uses" as a first-class front end, on par with its own app. DealGround is not alone in the broader category. Ascendix has published a Claude MCP integration for its AscendixRE CRM, and Terrakotta, which pivoted from an AI dialer to a contact-and-property database, now pushes results into Salesforce, HubSpot, Rethink and RealNex. But DealGround's is the most fully documented data-graph-as-server we have found.

For fifteen years, CRE software has competed on interface. Filters, maps, exports, dashboards. The vendor with the best screens won the seat.

MCP removes the screen from the equation. If a broker asks Claude a question and Claude answers from the vendor's records, the broker never sees the vendor's interface. What the broker experiences is the answer. Two things follow.

First, data quality becomes fully exposed. A pretty dashboard can hide a thin dataset because the broker only looks at what the dashboard chooses to show. An assistant querying the graph directly will surface every gap, every stale record, every unresolved LLC, because the broker asks whatever they want and the tool either has it or does not.

Second, the workflow layer stops being the vendor's moat. Which brings us to a retail landlord in Calabasas.

On August 27, 2026, NewMark Merrill Companies, which owns or manages more than 110 shopping centers across California, Colorado and Illinois, announced the completion of a twelve-week, six-session "AI Pioneer" program. Thirteen employees drawn from accounting, collections, property management, leasing, development, acquisitions, IT, marketing and HR, none of them technologists, built 53 working AI tools around bottlenecks in their own jobs. The company documented 3,783 hours of recovered annual capacity. Underwriting document review went from eight hours to under one. A day-long CAM reconciliation now runs in minutes.

Those numbers are self-reported by the company and its training partner, so read them as a case study rather than a benchmark. The direction is the point. Narrow workflow automation, the kind a mid-sized tenant-rep shop would have paid a SaaS vendor for two years ago, is now something the people closest to the friction can build themselves in a quarter.

What they cannot build is the data. NewMark Merrill's 53 tools all ran on records the company already owned: its own leases, its own security reports, its own reconciliations. A tenant-rep broker's binding constraint is the opposite. The records that matter, who occupies which suite, when their lease clock runs, which company just pulled a tenant-improvement permit or filed a WARN notice or moved its registered agent, live outside the brokerage and are scattered across county, state and federal sources that were never designed to be joined.

So the build-versus-buy line moves: build the workflow and buy the data. And when the workflow layer is your assistant plus MCP, "buy the data" means buying a graph your assistant can query.

Three things, and any vendor pitch that skips them is selling the protocol instead of the product.

It does not fix coverage. DealGround's graph is property-first: title, sale history, marketed OMs. That is the right shape for an investment-sales broker asking "who owns the low-basis QSR boxes in my territory." A tenant-rep broker asks a different question: "which companies in my submarket will need space in the next 12 to 18 months, and who inside them makes that call." Property-side data answers the first question well and the second one barely at all. An MCP server on top of a property graph gives you a fluent conversation about the wrong population.

It does not fix resolution. Every CRE data vendor now advertises LLC unmasking. The honest version, which DealGround does to its credit, returns a confidence score and the source trail so the broker can finish the research when the first answer is murky. Ask any vendor what fraction of the entities in your submarket resolve to a named decision-maker with a working contact. If they cannot give you a number with a definition attached, the MCP connection is a faster way to get "unknown."

It does not fix the lease clock. The single most valuable field for a tenant-rep desk is the expiration date, and for the long tail of King County tenants in sub-5,000-square-foot suites under non-institutional landlords, that date exists in no vendor's database at any price. We wrote about why that gap is the real market in our post on what a commercial tenant-rep data platform actually is. MCP makes an existing dataset easier to reach. It cannot conjure a field nobody collected.

Five questions, in the order a working broker should ask them.

1. What is the unit of your graph? Properties or occupiers? If the answer is properties, ask how many named occupiers in your submarket carry a resolved decision-maker and a dated signal. That is the count that matters for tenant rep.
2. What tools does the server expose, and which ones write or spend? DealGround publishes this list. A vendor who will not is either not ready or does not want you to know how thin the surface is.
3. Does every tool return provenance? For each record, where did it come from, when was it fetched, and what is the confidence? An assistant will happily repeat a stale record with total fluency. Provenance is the only defense.
4. Do calls run under my permissions and my team's audit trail? If a colleague's assistant can create lists or spend credits in your account, that is a governance problem dressed as a feature.
5. What can I do with the answer outside your product? Saved lists that live only in the vendor's app are a soft lock-in. The value of MCP is that the assistant is yours. Make sure the outputs are too.

We are biased here and we should say so. BrokerHQ is building occupier-side data for Seattle tenant-rep brokers: the company, the space, the signal, the person. BrokerHQ does not currently offer an MCP connection to that data. Whether and when it does is a roadmap decision, not a promise, and we would rather ship a graph that answers the tenant-rep question correctly than expose a thin one to Claude early.

What we believe, and what this month's evidence supports, is that MCP collapses the CRE software stack into two layers: the assistant the broker already pays for, and the data the assistant can reach. Everything in between, the dashboards and the drip sequences and the tour-book builders, is now something a motivated associate can build in a quarter, as NewMark Merrill just demonstrated with 13 people and no engineers.

The vendors who survive that collapse will be the ones whose data holds up under unfiltered questioning. For tenant rep, nobody has that graph yet, including us. But the buying question has already changed, and brokers who ask the five questions above will figure out quickly who is selling a protocol and who is selling records.

The strongest case against this post: MCP adoption inside brokerages is close to zero today, and the broker who actually connects Claude to a data vendor in 2026 is an early adopter at a firm with an IT lead. Most Seattle tenant-rep desks still run on CoStar exports and a spreadsheet, and NewMark Merrill needed a paid twelve-week program and a CEO who personally sponsored it to get 13 people building. A two-broker shop will not replicate that. On that reading, the interface still matters for years, and vendors are right to keep investing in screens.

We think the counter-argument is correct about timing and wrong about direction. The broker does not need to set up MCP for it to matter. The vendor's next hire, the brokerage's next intern, and the AI assistant already embedded in Outlook and Chrome will do it for them. The question is which data they will be pointed at when it happens.

What Is a Commercial Tenant-Rep Data Platform? A 2026 Buyer's Guide