Evaluating AI

Claude, MCP, and commercial real estate

Connecting Claude to your document repository is genuinely useful and takes an afternoon. Here is exactly what it gives you, and the four things it structurally cannot.

August 2, 2026 · 8 min read

What MCP actually is

The Model Context Protocol is a standard way for an AI assistant to talk to an external system — a document repository, a database, a ticketing system. Instead of a bespoke integration per tool, a system publishes an MCP server describing what it can do, and any MCP-capable assistant can use it.

For commercial real estate this matters because Egnyte publishes one. So a Claude user can, without much ceremony, give Claude the ability to search and read the company document library. This works. It is not a trick or a demo. We use MCP ourselves.

We should be straightforward here: Sevrel is built on Claude. We are not going to argue the model is the problem, because we chose the model.

What you get

A capable assistant that can find and reason over your leases without anyone uploading anything. For an individual — a analyst, a broker, a principal who wants to interrogate a document set — this is a real productivity gain, available today, at a per-seat price.

Search and retrieval across a connected repository.
Strong reasoning over long, dense contracts.
Drafting, summarizing, and open-ended analysis well beyond real estate.
On business and enterprise plans, content excluded from model training by default, with SSO and admin controls.

The four things a connector cannot do

These are not gaps in Claude. They are consequences of the shape of the thing: a connector is a capability granted to a person, inside a chat product, that runs when that person types.

It cannot enforce a company policy about who may query what. The connector's reach is the reach of the credentials that configured it. Roles that differ by property, fund, or ownership group have nowhere to live.
It cannot produce the record. An append-only log of what was asked, what was retrieved, and by whom is an artifact an auditor, a lender, or an LP asks for. A chat transcript in an individual account is not that artifact.
It cannot act unprompted. There is no Tuesday 6am job that reads the portfolio, notices a lease expiring in ninety days, and tells someone. Everything waits for a person to ask.
It has no data model. It reads a rent roll as a table of text. It does not reconcile that table against a budget, attach the result to a property record, or track a critical date forward through time — because there are no property records and no critical dates, only documents and conversation.

Why the fourth one matters most

The first three are governance problems, and governance problems feel abstract until the day they are not. The fourth is a daily problem.

Commercial real estate documents are not just long, they are structurally weird. A rent roll is a spreadsheet somebody exported from a system that has been customized for fifteen years. A shared-cost reconciliation is arithmetic spread across a lease, an amendment, and a ledger that use three different names for the same tenant. Reading these correctly is less about model quality than about knowing what a base year is and where to look for the cap.

A general assistant approaches each of these fresh, from context, every time. Purpose-built software encodes the structure once — including parsing machine-generated documents deterministically, in code, before the model is involved at all — so the same question produces the same answer next month.

An honest recommendation

If you are one person or a small team with a straightforward document set and no external reporting obligation, connect Claude to your repository. It costs very little, it works, and you will learn a great deal about what you actually want.

Come back to purpose-built software when the questions get repetitive, when more than a few people need different access, when something needs to happen without being asked, or when someone outside your company starts asking how the answer was produced. Those are the boundaries — not model quality, and not whether it can read a PDF.

See it on your own documents

The fastest way to settle any of this is to run it against a lease you already know the answers to.

Request Access See how Sevrel compares →