Ask ChatGPT which commercial loans mature next year: Circlemark is now an official ChatGPT app

Key numbers. We track 178,645 commercial loans across New York City and Texas, 15,177 of them active and maturing within twelve months of today, and 4,323 past a maturity stated on the recorded instrument with no payoff on file. As of this month those figures answer questions inside ChatGPT with one click: Circlemark is an official app in ChatGPT's directory, reviewed and listed by OpenAI in October 2026.

The honest version of that sentence: nothing new was built. The Model Context Protocol server we shipped for Claude in August is the same server ChatGPT now reaches, and the listing removes the one step that kept most brokers out — turning on a developer mode most workspaces never expose. This post is the walkthrough: how to install it, what a broker actually asks, what each answer cites, and where we drew the lines so an AI assistant cannot turn a modelled date into a fact.

Install it

Three steps, about a minute, and it works with a free seat.

1. Install the app. Open Circlemark in the ChatGPT app directory and select Install plugin. No developer mode, no URL to paste.
2. Sign in and approve. ChatGPT sends you to circlemark.ai to sign in — it never asks for your password in the chat — and shows a consent screen listing what it may do under your account. Click Allow.
3. Ask. Start with a search; it is free. The assistant is told what every tool costs, so it says so before anything spends a credit.

In ChatGPT Work the same app runs multi-step research across a whole borough rather than one question at a time. In Codex, or in a workspace whose admin has not exposed the directory, the server is added by hand — Settings → Security and login → Developer mode, then the URL https://circlemark.ai/api/mcp — and the result is identical. The setup page has both routes and the admin modes (an Enterprise admin can allow the app read-only, which leaves search, lender rankings and the watchlist on and contact reveals off).

The walkthrough

The two-minute recording on the setup page is the session below, made on September 12, 2026 against the live database. The counts are that day's; the record moves every week, so yours will differ.

First question. "Which New York City commercial loans between $4M and $30M have already passed their maturity date with no payoff recorded?"

ChatGPT calls search_loans for the New York metro with a maturity window from twelve months ago to today and the amount band, and reports 134 loans, property by property: address, borough, amount, the maturity date, the lender as recorded, the borrower entity and the recording number. Then it adds, unprompted, the sentence we care most about: no payoff recorded does not mean the borrower is delinquent — it means the public record still shows an active lien after the stated maturity, and the borrower may have an unrecorded extension or a refinance elsewhere. That caveat is not the assistant being cautious; it is what the tool's own description says the row means.

Second question. "Which of those maturity dates come from the recorded document, and which are estimates?"

All 134 are stated, and ChatGPT explains why: a maturity window set entirely in the past matches only dates read from a document. An estimate that has lapsed says the model's assumed term was wrong, not that the loan is late, so it never enters a past-due answer. This is the same rule as the search page's past window and the methodology; the point of the connector is that the rule survives the trip into a chat.

The shape of the session: the facts come from the recorder through tool calls, the reasoning happens in the chat, and anything that costs a credit asks first.

Third question. "Narrow it to Brooklyn, debt-fund lenders only."

Four loans, each with its lender, stated maturity and recording number, and ChatGPT's own ranking of which to call first. The ranking is the assistant's reasoning, not ours — we hold no opinion on which borrower to call — but every fact in it is a field from the record.

Fourth question. "Who is behind the borrower on the first one? Reveal its contact."

Before calling reveal_contacts the assistant states the cost — one credit, or free if the property was already unlocked on your seat — and waits. The answer names the people the public record ties to the borrowing entity, each row citing where the name came from: the HPD registration, the recorded instrument's signature page, the assessor roll, the Texas Comptroller's officer filing. Work emails come from vendor lookups and say so. A direct phone number is a separate, explicit request (find_phone, five credits, charged only when a number lands), and the assistant will not run it unasked.

What each tool does, and what it costs

Credits are the allowance your seat already has — 10 for the life of a free seat, 100 a month on Data, 300 on Pro — and nothing is sold inside ChatGPT. The lender tools are revealed appetite from recorded liens: who actually closed, dated to the last deal. They carry no quoted credit box and no originator's name, because we hold neither, and the assistant is told not to invent either.

Where we drew the lines

Putting a paid dataset behind a language model invites exactly the failures our methodology exists to prevent, so the server enforces three rules whatever assistant is on the other end.

A modelled maturity never travels as a fact. ACRIS has no maturity field and the Texas clerk's index has neither amount nor date; every date anywhere was read off the instrument, read from a securitization filing, or modelled. Each row carries its basis — stated, reset for a first rate-change date read from the recorded rider on 30-year paper, stated_term where the document states a term but prints no date, or the assumed term — and the tool description instructs the assistant to say which. The ≈ survives the trip.

Spending is explicit. Searches are metered like the web app. Anything that costs a credit says so in its own name and description, so the assistant asks before unlocking, and a phone hold is refunded when no number is found.

ChatGPT never sees credentials or the database. The connection is OAuth 2.1: you sign in on circlemark.ai, approve the scopes, and the assistant holds a revocable token for your seat — the same identity, plan and allowances as your login. Nothing is shared across a workspace unless its admin chooses to.

The by-hand route

Everything above can be done without us, and a tutorial should say how. A past-maturity list for New York starts in ACRIS: search by document type for mortgages and agreements in your amount band, open each document image, and read the maturity off the page — it is stated on about half of plain mortgages and almost never on a consolidation agreement, which is why so many dates in the market are estimates to begin with. Then search the same lot for a satisfaction or an assignment to see whether the lien is still open, and read the HPD registration for the people. The by-address lookup tutorial walks one building through it, and the maturing-loans guide covers the three sources every list is built from. The app is those steps, run for every loan at once, with the labels kept.

The moment this is for

MBA counts 17% of all outstanding commercial and multifamily balances maturing in 2026, and S&P expects the wall to peak in 2027. Most of those decisions — refinance, extend, inject, sell — happen at buildings nobody writes about, financed by banks with no servicer disclosure, visible only in the county record. If an assistant can shortlist the loans, say which dates it is sure of, name the decision-maker with a citation and draft the first email in one sitting, the cost of starting drops to a question — and that is the only thing that changed this month.

Prefer Claude? The same connector answers there, and the Claude tutorial in this series walks the same questions.