Ask Claude which CRE loans mature next year: Circlemark now provides an MCP connector

Commercial mortgage maturity data now answers questions in plain English. How we put over 100,000 NYC loans inside Claude via the Model Context Protocol, what an AI assistant can and can't do with recorded-document data, and why the connector never presents a modelled maturity as a fact.

By Circlemark · Aug 30, 2026 · 5 min read · Figures as of Aug 31, 2026

Key numbers. We track 108,989 commercial loans across the four ACRIS boroughs, 13,240 of them active and maturing within twelve months of today — against a national backdrop of $875 billion of commercial and multifamily debt coming due in 2026. As of this week, that dataset answers questions in plain English, inside Claude: Circlemark now provides a Model Context Protocol connector.

That sentence deserves immediate deflation. There is no new AI in our product. What shipped is a door: an MCP server that lets an assistant you already use call the same search and reveal machinery the web app uses, under your account, with every provenance rule intact. This post is about what that's actually good for, where we drew the lines, and how to connect.

What a broker can ask

MCP — the Model Context Protocol — is an open standard for giving AI assistants tools. Connect Circlemark, and Claude gains four:

Tool	What it does	Cost
`search_loans`	Search loans by maturity window, borough or county, size, lender, LTV, distress signals, borrower or street	Free, within your plan's daily limits
`reveal_contacts`	Unlock the people behind a property's loan — principals from the public record, each citing its source, plus work emails	1 credit (free if already unlocked)
`find_phone`	Run the direct-phone lanes for one named person	5 credits, charged only when a number lands
`get_credit_balance`	This month's allowance and usage	Free

Which turns a workflow that used to be filters-and-columns into a conversation:

You: Which multifamily loans in Brooklyn over $2M mature in the next 12 months? Sort by size.

Claude (calls search_loans with county: ["nyc-kings"], due: "next12m", amountMin: 2000000): There are 214 matches. The largest is a $28.5M loan from 2016 on a 96-unit building in Crown Heights — the maturity is modelled (≈ 10-yr bank term, nothing stated on the recorded document). Three of the top ten carry stated maturities read from the instrument itself…

You: Who's behind the Crown Heights one?

Claude: Revealing contacts costs 1 credit from your monthly allowance — you have 287 left. Proceed?

Stylized chat: a broker asks which Brooklyn multifamily loans over $2M mature within 12 months; the assistant calls Circlemark's search_loans tool, reports 214 matches led by a $28.5M Crown Heights loan with a modelled ≈10-year maturity, and asks before spending a credit to reveal the owner The shape of the workflow: facts come from the recorder via tool calls, the reasoning happens in the chat, and anything that costs a credit asks first.

The assistant composes what the app never could: "…and draft a first-touch email that mentions their 2016 rate," or "which of these borrowers shows up on more than one maturing loan?" — the reasoning happens in the chat, the facts come from the recorder.

Where we drew the lines

Putting a paid dataset behind a language model raises exactly the failure modes our methodology exists to prevent, so the connector enforces three rules.

A modelled maturity never travels as a fact. ACRIS has no maturity field — every NYC maturity date anywhere is stated on the recorded document, read from a securitization filing, or modelled. The connector returns each date with its basis (stated or the assumed term), and the tool description instructs the assistant to present estimates as estimates. The ≈ marker survives the trip into the chat.

Spending is explicit. Searches are metered like the web app. Anything that costs credits says so in the tool's own name and description, so the assistant asks before unlocking. Phone lookups never run unasked — a number is a separate, per-person request, and the 5-credit hold settles only if a number actually lands.

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

Why a connector beats pasting exports into a chat

You could always download a CSV and drop it into an AI chat. Three things break when you do: the data is stale the moment it's exported; the provenance is stripped (a modelled date and a stated one look identical in a spreadsheet); and the export ceiling means the model reasons over a sample, not the market. A connector inverts this — the assistant queries the live projection, page by page, with the labels attached, and a week later the same question gets this week's answer. Recorded documents change slowly; lifecycle events don't. 196,763 lifecycle events are matched into the dataset so a satisfied or consolidated loan never surfaces as a lead, in the chat or anywhere else.

The moment this is for

The maturity wall is not abstract: MBA counts 17% of all outstanding commercial and multifamily balances maturing this year, and S&P expects the wall to peak in 2027. Most of those decisions — refinance, extend, inject, sell — happen at buildings nobody writes articles about, financed by banks with no servicer disclosure, visible only in the county record. The brokers winning that market are the ones who reach the right owner first. If an assistant can shortlist the loans, name the decision-maker with a citation, and draft the outreach in one sitting, the cost of starting drops to a question.

Connect it

Takes about a minute: Settings → Connectors → Add custom connector → https://circlemark.ai/api/mcp, sign in, approve. The step-by-step — including Cowork workspaces and what each tool costs — lives at circlemark.ai/claude. The connector works with any Circlemark account, including free seats with their 10 monthly credits.

New York City is live today; Dallas–Fort Worth counties are ingesting now and answer through the same tools as their recorder backfills complete.

Frequently asked questions

What is an MCP connector?: The Model Context Protocol is an open standard that lets an AI assistant call tools on an external service — with your permission, under your account. A connector is an MCP server the assistant talks to. Circlemark's connector exposes loan-maturity search and owner-contact reveals to Claude, so the assistant queries the live database instead of guessing from training data.
How do I connect Circlemark to Claude?: In Claude (web, desktop, or a Cowork workspace): Settings → Connectors → Add custom connector → enter https://circlemark.ai/api/mcp. Claude registers itself, sends you to circlemark.ai to sign in, and asks for your consent. Setup takes about a minute; see circlemark.ai/claude for the step-by-step.
Does Claude get access to my Circlemark account or data?: Only what you approve. The connection uses OAuth 2.1: you sign in on circlemark.ai — never in the chat — and Claude receives a token scoped to your seat. Searches and reveals run as you, under your plan's allowances, and you can revoke the connection at any time from your account.
What does it cost to use Circlemark inside Claude?: The same as the web app — the connector is another door to the same account. Searches are metered by the plan's daily fair-use limits. Unlocking the people behind a property costs 1 reveal credit (free plans include 10 per month, paid seats 300), and a direct phone number is 5 credits, charged only when a number is found. The assistant is told each tool's cost up front, so it asks before spending.
Will the assistant present an estimated maturity date as a fact?: The data won't let it. Every maturity the connector returns is labelled: a stated date read from the recorded document, or a modelled estimate with the assumed term attached. The tool's own description instructs the assistant to present estimates as estimates — the same ≈ rule the app follows.
Which AI assistants work with the Circlemark connector?: MCP is an open standard. We test against Claude — claude.ai, Claude Desktop, and Cowork workspaces — which is where the connector is verified today. Other MCP-capable assistants can speak the same protocol; support for them is on the roadmap as we verify each one.