Crunchy Sheets 2.0: QuickBooks, Custom Commands, and a Real Audit Trail

6 min read

Crunchy Sheets launched in February 2026. Seven weeks later, this is the first major release, and there's enough in it to warrant a proper write-up. If you already have the add-on, you have all of this automatically — no reinstall.

TL;DR: Now officially linked to QuickBooks Online — can pull GL data directly into your financial model; User-defined commands (i.e., like "skills"); Proves its work.

Here's what shipped.

QuickBooks Online lives in the sidebar

You can now pull structured reports from QBO directly into the workbook you're already in. Two ways to do it.

Click the QB button and a form opens: Profit & Loss, Balance Sheet, Trial Balance, Sales by Customer, transaction detail exports. Pick a period, pick accrual or cash, pick your columns, and the report lands in the Sheet with the hierarchy intact and the subtotals tied.

Or just ask for it in plain English. "Pull last quarter's P&L by month." The assistant routes through the QBO tools and writes the result into the Sheet with the same renderer. Same output, fewer clicks.

A few things worth flagging. Transaction-detail exports render line by line and let you pin a preferred column order. Each user gets their own isolated QBO key at connect time — disconnect actually revokes the refresh token at Intuit, no residue. And the connection does not require any new Google OAuth scopes, which means the install story in the Workspace Marketplace hasn't changed.

This single feature is why I'd call this a 2.0. If you close books in QuickBooks and present in Sheets, the export loop is the part of your month that is least fun and most error-prone. It is now optional.

Custom commands

You can now define your own slash commands. If you run the same variance check every month, or have a recurring reconciliation routine, or want a formatting pass that matches your firm's house style — write the instruction once, name it, and call it with a slash.

Commands can be created in plain English ("make a command that builds a month-over-month variance tab and flags anything more than 10% off") or through the /commands palette. A custom command can also fork a built-in skill, so you inherit everything the existing skill already knows and add your own constraints on top.

This is the feature that turns the assistant into something that fits your workflow, rather than something you adapt to.

It pays attention to what you highlight

Selecting a range now scopes the request. Highlight B2:D20 and ask "clean this up" and the assistant only touches those cells. Highlight a block of formulas and ask "what's driving this" and it traces just that block. This sounds small. It stops being small the first time you're in a 40-tab workbook and want the assistant to stay in one place.

Formula X-Ray and Prove It

Two features for people who have to show their work.

/xray B14 returns a color-coded breakdown of the formula in B14, walks up to three levels of dependency (capped at 100 cells), and ties the rebuilt formula back to the value actually sitting in the cell. It's read-only — it will not propose changes. It's for when you opened someone else's model and you need to understand what a cell is actually doing.

/prove goes further. It builds a standalone proof tab that traces a number back to its source cells in a form another reviewer can read end-to-end. If you're handing a model to a board, a lender, or an auditor, this is the tab you point them at.

Both of these exist because the honest answer to "can I trust this number?" is usually "let me rebuild it myself," and that is the work the assistant should be doing for you.

New commands, new actions, one rename

Other commands that landed since launch: /format for a full or targeted formatting pass against the financial-model color conventions, /explain for a structured walkthrough of the active tab, /audit for safe tab-deletion recommendations based on the cross-reference graph, and /rescan when you've made structural changes and want the cached model rebuilt.

Under the hood, the action system gained a run_script type, which lets the assistant execute short Apps Script snippets for bulk operations — sandboxed, with Gmail, Drive, and external-request APIs blocked. There are also new actions for borders, column widths, notes, named ranges, tab colors, sheet moves, and freeze rows.

One rename: the Tab Audit skill is now "Find & Delete Unused Tabs," and it's stricter than it was. It only proposes deleting tabs that are provably safe — empty, isolated, or hidden-and-disconnected — and it warns rather than deletes when a hidden tab is still referenced somewhere.

Under the hood

The stuff you wouldn't notice if it works and would notice immediately if it didn't.

A lightweight fast path for mechanical requests ("create a tab called Q2") that skips the full workbook model entirely. A cached structural summary — roughly 2–5K tokens instead of the full workbook — that the assistant uses for most requests and only expands when it actually needs more data. Client-side caching that lifted the cache ceiling on large workbooks. A hybrid cross-reference loader that removed a 15–33 second lag on the structural path. Hidden sheets now tracked as metadata stubs so the assistant doesn't lose track of them.

On the reliability side: a stop button, a real 75-second timeout, a silent-hang safety net, truncated-JSON recovery, proactive staleness checks before each request, and a _CrunchyLog sheet that records every action the assistant applies so you can audit what it did after the fact. Formula X-Ray is gated against hallucinated responses. Prove It is reliably invocable from suggestion chips and compound requests. Apply All has real error recovery.

The Sonnet tier was upgraded from Claude Sonnet 4.5 to Claude Sonnet 4.6. Opus-tier skills (cash flow forecasting, unit economics, cohort analysis, scenario modeling, Prove It) continue on Claude Opus 4.6.

One for developers

There's a new Model Context Protocol endpoint at /api/mcp. If you want to query the same tools from a different client — Claude Desktop, Cursor, Zed — point your MCP client at that URL. Same backend as the sidebar, different front door.


That's the release. If you already use Crunchy Sheets, the update is live for you now. If you don't, it's on the Google Workspace Marketplace — installs in about a minute and lives in the sidebar of any spreadsheet. Future release posts will show up here when there's something worth saying.

Subscribe for thoughtful, cutting-edge rants from Gary

and the opportunity to beta-test new finance superpowers and parenting superpowers