Stripe, Square and Ramp join NumbersGame's QuickBooks and Xero MCP

NumbersGame, a vendor that wires AI assistants into accounting data, has added Stripe, Square and Ramp to the MCP server it offers alongside QuickBooks Online and Xero. The change matters to accounting firms that work in Claude or ChatGPT, because those assistants can now reach payment and spend records in the same conversation as the books. Here is what the update covers, what it is good for, and where we would slow down.

MCP in plain terms

MCP stands for Model Context Protocol. It is an open standard that lets an AI assistant pull live data from a connected service instead of working only from what you paste into the chat. Claude or ChatGPT acts as the client. The vendor runs the server side, which talks to QuickBooks Online or Xero and returns records the assistant can read and reason about.

Until now that server covered the two bookkeeping platforms. The update adds three more sources, so payment and spend data can sit beside the ledger data in one place.

The three connectors at a glance

Stripe and Square sit on the payments side. Stripe handles card payments for online businesses, while Square is common with retail, hospitality and service clients that take payments at the counter. Ramp covers corporate cards and spend management.

The practical effect is that questions spanning systems no longer need exports. The assistant can look at processor records and ledger records in the same turn, which is where much of the manual copy-paste used to live.

What can a firm actually do with it?

In principle, the sorts of questions that used to need three open tabs:

• Compare Stripe payouts for a month against the deposits recorded in QuickBooks Online, and flag gaps.
• Check Square fees against recorded sales for a cafe client before the close.
• Review Ramp card transactions and ask which ones lack receipts or sit in odd categories.

Treat these as drafts, not answers. The assistant can misread a field or miss a payout, so someone on the team still checks the numbers before anything reaches a client.

Before you connect client data

A connector that reads financial records deserves the same care as any other integration. A few checks we would run first:

• Confirm whether the connection is read-only or can post entries, and set it accordingly.
• Make sure the client knows a third-party tool is reading their data, and that the engagement terms allow it.
• Limit which team members can authorise connections, especially in ChatGPT, where connector availability depends on the plan.
• Start with one client and one narrow question rather than switching the whole portfolio at once.

The risks worth naming

An assistant that answers across Stripe, Square, Ramp and the ledger is only as reliable as the data behind it and the prompt in front of it. Confident wrong totals are the failure mode to expect: the answer reads well and the figure is off. There is also scope creep. A tool bought to answer questions can drift into posting transactions, and unreviewed entries are how books quietly rot.

None of this argues against the update. It argues for treating the assistant as a junior team member whose work gets reviewed, not as a source of truth.

How should a firm start?

Pick one client with clean books and a single, checkable question, such as comparing last month’s Square deposits to the ledger. Run it, then verify the answer by hand, line by line, once. If the assistant gets it right and the time saved is real, write down the workflow and decide which other clients it suits. If it gets it wrong, you have learned that cheaply, on one client, with a small blast radius.