The short answer: most owners need no server at all.
As of September 2026 there are three ways to let an AI work with your QuickBooks through MCP, the protocol AI assistants use to reach other software. Intuit's own connector inside Claude, which takes a few clicks. Intuit's open-source QuickBooks Online MCP server, which a developer runs. Or a custom server, when QuickBooks is one of several systems the AI has to see at once. If you own the business and want answers from your books, start with the first one.
The rest of this page is what each option can touch, who it suits, and the settings worth choosing before an AI gets write access to your accounting. If MCP itself is new to you, the plain-English guide to MCP servers covers the idea in ten minutes.
The QuickBooks connector in Claude
In Claude, open Connectors, search for QuickBooks, connect, and sign in with your Intuit login. That's the setup. According to Intuit's help article, it can generate a P&L and cash-flow reports, benchmark you against similar businesses, import and categorize transactions, create and send invoices and estimates, create customers and items, and make payment links.
Three details from the same page matter more than the feature list. Destructive actions ask for your confirmation before QuickBooks changes anything. Intuit says your QuickBooks data isn't used to train Claude. And the connector is available to US customers only for now. Intuit's July 28 announcement describes the same invoicing, estimate, payment-link, and payroll-lookup actions arriving in ChatGPT as well.
Who it suits: an owner or office manager who wants to ask “which customers are more than 30 days late?” and get an answer from the live books, or draft this week's invoices by describing them. Nothing to host and nothing to maintain.
The QuickBooks Online MCP server on GitHub
Intuit also publishes quickbooks-online-mcp-server, an Apache-licensed server a developer runs on their own machine. Its README lists create, read, update, delete, and search across 29 QuickBooks entities (customers, invoices, bills, vendors, payments, purchase orders, time activities, and more) plus 11 financial reports.
It is not a click-to-connect product. You register an app on Intuit's developer portal and complete a one-time OAuth sign-in. A sandbox company can use a localhost callback, but production needs a public HTTPS callback for that first authorization, and the refresh token has a 100-day window. The useful safety feature is three switches: QUICKBOOKS_DISABLE_WRITE, DISABLE_UPDATE, and DISABLE_DELETE. Set them and those tools never load, leaving the AI with lookups and reports only.
Who it suits: a team with a developer who wants QuickBooks inside their own agent or internal tool rather than inside Claude, or who needs entities the connector doesn't cover yet, like bills and vendor payments.
When the answer lives in more than your books
Both Intuit options see QuickBooks and nothing else. Plenty of the questions worth asking cross systems: which jobs finished last week and still haven't been invoiced, or which customers with open estimates also have an overdue balance. The job system knows half of it; QuickBooks knows the other half.
That's where a custom MCP server earns its cost. It sits on your infrastructure, reaches QuickBooks and the other systems through their APIs, and exposes only the actions you've approved, each marked read, write, or destructive. I build these, and the permission layer I use is open source: mcp-gatehouse adds approval gates and an append-only log of every call, so “what did the AI change?” has an answer months later. The reverse-logistics case study shows one running in production.
Who it suits: a business past the point where one system tells the whole story, with a leak big enough to price. If you can't put a dollar figure on the cross-system question yet, Option 1 is the better spend.
The three options, on the questions that decide it
| Claude connector | Intuit's server | Custom server | |
|---|---|---|---|
| Who sets it up | You, in a few clicks | A developer | Whoever builds it |
| What it reaches | QuickBooks only | QuickBooks only | QuickBooks plus your other systems |
| Where it runs | Intuit's side | A machine you control | Your infrastructure |
| Write access | Yes; destructive actions ask you first | Yes, unless you switch writes, updates, or deletes off | Whatever you decide, per tool |
| Record of what the AI did | Your QuickBooks history | Whatever you build | A log you own, if it's built in |
| What you need | A QuickBooks subscription and Claude | An Intuit developer app and a developer | A scoped build |
Checked against Intuit's help article, announcement, and GitHub README in September 2026. Connectors change fast; confirm on those pages before you decide.
An AI that can send invoices can send the wrong one.
Every option above can write to your books. That's the point, and it's also the risk: a wrong customer, a duplicated invoice, a payment link sent twice. Five settings worth choosing on day one, whichever route you take:
- 01Try it on a sandbox or a copy of your company file before the real books.
- 02Start read-only. Reports and lookups answer most of the questions owners actually ask.
- 03Connect with your own login, not a shared admin account, so every change traces to a person.
- 04Decide which actions need a human yes. Sending an invoice to a customer is the obvious one.
- 05Tell your bookkeeper or accountant before an AI starts creating entries they'll reconcile.
And the honest skip: if nobody in the business uses an AI assistant yet, none of this is urgent. QuickBooks works exactly as it did before. Connect it the week someone asks a question the reports can't answer fast enough.
Keep reading: MCP vs. API: what's actually different? Or, if the books are where the hours go, what to automate first.