QuickBooks MCP: How to Connect QuickBooks to Claude and Automate Your Close

If you run accounting at a growing company on QuickBooks Online, you've probably noticed something frustrating about AI tools. 

They can do almost anything — summarize documents, draft emails, write SQL — but they can't actually touch your accounting data. Ask Claude to pull your balances or tie out a reconciliation, and the answer is always the same: "You'll need to export that from QuickBooks first."

MCP changes that. A QuickBooks MCP server lets AI assistants read and act on your QuickBooks data directly — no exports, no copy-paste. 

This guide covers what a QuickBooks MCP server does, how to get one connected, and how pairing it with Numeric's MCP turns Claude into a teammate that can help run your month-end close.

What is a QuickBooks MCP Server, and Why Should Accountants Care?

MCP stands for Model Context Protocol — an open standard introduced by Anthropic that lets AI assistants connect to outside tools and data sources through a common interface. 

Think of it as a universal adapter. Instead of AI that only works with what you paste into a chat window, MCP lets Claude operate inside the systems you already use — QuickBooks, Google Sheets, Slack, Gmail, Numeric, and so on.

A QuickBooks MCP server exposes your QuickBooks Online data and actions to an AI client. Depending on which server you run, that includes:

• Reading your chart of accounts, customers, vendors, items, and classes
• Pulling transaction data — invoices, bills, journal entries, payments, deposits
• Generating reports — P&L, balance sheet, cash flow, A/R aging, A/P aging
• Creating or updating transactions when you grant write permission
• Querying your GL in plain English without building an advanced report

The problem this solves is one accounting teams know intimately. Your GL lives in QuickBooks. Your close checklist lives in a spreadsheet or a project tool. Your reconciliations live in Excel or Google Sheets. Your status updates live in Slack or email. Every workflow that crosses more than one of these systems requires a human to manually carry information between them.

MCP collapses that gap. And for teams on QuickBooks, it's the first time AI can actually do useful work on your books without you building an integration from scratch.

What You Can Do Once QuickBooks is Connected

You're not moving all of your QuickBooks work into Claude. You're using Claude as a layer on top of QuickBooks that lets you ask questions, pull data, and string actions together faster than clicking through the UI.

A few examples of what controllers and accounting managers do on day one:

Query the GL in plain English. "What did we spend on AWS in Q3?" Claude builds the report, filters to the relevant account, pivots by vendor, and returns the answer. Same question in QuickBooks alone would take four to six clicks, a saved filter, and an Excel export.

Pull reports without hunting for them. "Give me our trailing 6-month P&L by month, comparing this year to last year." Claude runs it, formats it, and hands you a clean output you can drop into a board deck.

Investigate anomalies at the transaction level. "Why did travel expense jump in October?" Claude pulls the transactions, groups by vendor, and flags the one-time charges.

Reconcile against external data. Upload a vendor statement or a bank report, and Claude compares it to what's posted in QuickBooks and flags the differences.

Draft journal entries for review. Give Claude the facts — "We prepaid $48,000 for a 12-month software subscription on September 1" — and it drafts the entry, including the amortization schedule. You review, approve, and post.

Every question becomes a conversation instead of a scavenger hunt.

How to Set Up QuickBooks MCP: Three Paths

Three main paths exist today. The right one depends on how technical your team is, whether you want a hosted or local setup, and whether you're working from raw GL data or on top of a platform that's already organized your close.

Path 1: Intuit's Official MCP Server

In October 2025, Intuit released an official QuickBooks Online MCP server as an early preview. It's an open-source TypeScript/Node.js package built on the QuickBooks Online Accounting API.

Two things to know:

1. It runs locally via stdio transport. That means it works with Claude Desktop and similar clients that support local MCP servers — not with the web versions of Claude.ai or ChatGPT.
2. The preview is aimed at developer experimentation. The server supports both sandbox and production environments through the QUICKBOOKS_ENVIRONMENT config variable, but Intuit has positioned this release as a place for developers to test the API before deploying their own apps, not as a turnkey connection to live books.

Setup:

bash
# Requires Node.js 18 or higher
git clone https://github.com/intuit/quickbooks-online-mcp-server
cd quickbooks-online-mcp-server
npm install

You then register an app in the Intuit Developer Portal, grab your client ID and client secret, set the QUICKBOOKS_ENVIRONMENT variable to sandbox or production, and either supply an existing refresh token and realm ID or run the built-in OAuth flow to generate them. 

Configure your MCP client (Claude Desktop, Windsurf, etc.) to point at the server and you're connected.

Use this path if you want to understand how QuickBooks MCP works under the hood or build a custom internal tool. Skip it if you just want to connect live books and start using Claude for close work.

Path 2: Third-Party Hosted or Open-Source Servers

Because Intuit's server is local-only, the community has built alternatives that fill different gaps:

• Hosted bundles like MCPBundles and Pipedream offer QuickBooks MCP servers accessible over HTTP, so they work with any MCP client — Claude.ai, Claude Desktop, Cursor, n8n — not just desktop ones. MCPBundles, for example, wraps the full QuickBooks Online API with financial reports, batch operations, and a SQL-like query tool.
• Community GitHub projects like laf-rge/quickbooks-mcp target financial professionals working with production books and include production-oriented features like AWS Secrets Manager for credential storage.
• Framework-specific packages like the Laravel QuickBooks MCP let engineering teams embed a server into their own application stack.

The tradeoff is the same one you make with any third-party API integration: you're trusting a vendor or open-source maintainer with credentialed access to your books. Do the standard due diligence on who's maintaining it, what permission scopes it requests, how tokens are stored, and whether there's an audit trail.

For most accounting teams, this is more flexibility than you need. Which brings us to the third option.

Path 3: Connect Through an Accounting Platform That Already Has MCP

This is the path most useful for teams doing real close work. Instead of connecting QuickBooks directly and building on raw GL data, you connect through Numeric that already sits on top of QuickBooks and organizes your close, reconciliations, flux, and reporting.

The distinction: a direct QuickBooks MCP gives Claude access to your GL. A Numeric MCP gives Claude access to your GL plus your close checklist, reconciliations, flux explanations, task assignments, and review history. All already structured, already tied to the workflows your team runs every month.

Claude.ai or Claude Desktop:

1. Go to Settings → Connectors
2. Add the Numeric MCP server using https://api.numeric.io/mcp
3. Log in with your Numeric credentials when prompted
4. Start a conversation.

Claude Code:

bash
claude mcp add --transport http numeric https://api.numeric.io/mcp

Then run /mcp to authenticate.

Once connected, Claude has the same access as your Numeric user account. Your QuickBooks data flows through Numeric's existing sync — already reconciled, already categorized, already mapped to your close workflows.

Why Pairing a QuickBooks MCP With Numeric MCP Changes the Math

A direct QuickBooks MCP is useful, but it has a ceiling. It can query your GL and draft entries, but it doesn't know anything about your close process, flux commentary, reconciliation status, or which preparer owns which task. 

Claude can answer "What did we spend on software last month?" in seconds but can't answer "Is the software expense reconciliation done for October, and who's the reviewer?" — because that context doesn't live in QuickBooks. It lives in how your team actually closes the books.

Numeric's MCP server exposes 20+ tools across three domains:

1. Workspace and context. Users, open and closed periods, entities, GL connections, holidays, tag groups, permissions.
2. Close task management. List, create, assign, comment on, and submit tasks; pull event history for audit trails.
3. Financial reporting and GL data. Pull saved reports, build ad-hoc statements, query transaction lines, read and write flux explanations.

Because Numeric already syncs with QuickBooks Online (along with NetSuite, Xero, and Sage Intacct), the Numeric MCP gives Claude the ability to reason over both the raw financial data and the operational context of your close.

You can ask things like:

• "Which reconciliation tasks are overdue, and for each one, show me the current account balance."
• "Pull last month's flux explanation on Travel & Entertainment, then show me the underlying transactions so I can verify it."
• "Draft a board-ready P&L summary for October, with commentary pulled from our flux tasks in Numeric."
• "Every account that moved more than 15% month-over-month — list it, pull the transactions, and draft a variance explanation I can review."

Multi-system workflows that used to require a human stitching data across tools become a single prompt.

Practical Accounting Workflows to Try First

Five starter workflows that show what's actually possible once you're connected. Each replaces an hour or more of manual work, and none requires code.

1. The morning close standup. "Check all open close tasks for the current period. List anything overdue or due today. For each task, pull the current balance for the related account. Post a summary to our #accounting-close Slack channel." Replaces the 20 minutes of list-pulling, balance-checking, and status-drafting that starts every close morning.

2. The AWS question. "What did we spend on AWS in Q3? Break it down by month, and flag any month more than 15% above the Q3 average. For each flagged month, pull the underlying transactions." Three to four minutes of conversation versus 20 to 30 minutes of clicking through a filtered expense report and building a pivot.

3. The flux narrative. "Pull our month-over-month income statement from Numeric. For every line that changed by more than $10,000 or 10%, pull the underlying transactions, identify the drivers, and draft a one-sentence variance explanation. Write the explanations back to Numeric for my review." Replaces the first draft of flux analysis — for many teams, the single most time-consuming part of close.

4. The ad-hoc board question. Your CFO emails at 4 p.m. on a Tuesday: "What's our trailing 6-month gross margin trend, and what's driving it?" "Build a trailing 6-month income statement, calculate gross margin by month, and identify the three accounts with the largest impact on the change. Pull the flux commentary from Numeric, or if there isn't any, draft an explanation." Three minutes of review instead of 45 minutes in your GL and Excel.

5. The audit packet. "For the October close, pull the full task history: who prepared each reconciliation, who reviewed it, when it was submitted and approved, and any comments. Flag any task where the same person submitted and approved. Export it as an Excel file for the auditors." Replaces half a day of pulling activity logs for PBC requests.

Security, Permissions, and What to Tell Your Auditor

Controllers reading this are, correctly, thinking about security. Three things to be clear on.

MCP uses standard OAuth. When you connect QuickBooks — directly, through a third party, or through Numeric — you go through Intuit's OAuth flow. Your password never touches the MCP server or the AI client. You can revoke access from your QuickBooks admin panel at any time.

Permissions mirror the connected user. The AI client has the same access as the account it connects through. If you don't want Claude to draft journal entries for review/posting with write permissions, connect through a read-only user. If you want writes only in a test environment, connect your sandbox company first.

Every action is logged. Numeric logs every tool call Claude makes — which tool, which parameters, which user initiated the session, and the result. The MCP becomes part of your audit trail.

The right starting posture for most teams: connect with read access, use it for a month, prove out the value, then scope up to write access for specific workflows where the time savings justify it.

QuickBooks MCP: Where This Goes Next

The QuickBooks MCP story is moving fast. Intuit has shipped an early preview, the community has built production-ready alternatives, and accounting teams are figuring out which workflows are genuinely better when AI can read the GL.

For teams running close on QuickBooks today, the move is to get connected, start with read-only workflows, and build from there. The close has always demanded both precision and speed. MCP doesn't change what precision requires — but it does change what speed looks like.

If you're curious how Numeric layers on top of QuickBooks to make this work, schedule a demo.

FAQ