Ask instead of building the report

Most questions about a business die somewhere between asking and answering. "Which products lost money last month?" is a ten-second question and a forty-minute export. So you don't ask it, and you run the month on the two numbers you already have on screen.

Panelque now speaks the Model Context Protocol – the standard AI assistants use to read data out of an app. Connect one and the question stays ten seconds.

It reads your costs, not just your sales

This is the part worth being precise about. An assistant wired to Shopify can see orders and revenue. It cannot see what a candle costs you to pour, because nobody told Shopify.

Panelque did tell it. So the connection hands over the cost side: the bill of materials behind each product, the margin on every variant, stock on hand and what it's worth, what's about to run out, operating expenses by category, refunds and what those orders still cost you, gateway fees, per-order breakdowns down to the individual cost line.

"What is sitting in my warehouse?" becomes a question with an answer.

Connecting one

There are two ways in, and which one you use is the assistant's decision rather than yours.

Claude Code and Claude Desktop read a config file, so they take a token. Settings → AI access, name it after the thing you're connecting – "Claude on my laptop" – and create it. The page hands you the install command or the config block with the token already in it. Paste, restart the assistant, done.

The token is shown once. We store a SHA-256 hash of it and nothing else, so if you come back later the install example is still there but with a YOUR_TOKEN placeholder where yours used to be. That isn't us being coy – we genuinely can't print it again.

ChatGPT and Claude.ai on the web have nowhere to paste one. Their connector forms offer OAuth or nothing, which for a long time meant they simply couldn't reach Panelque at all. Now they sign in instead: add panelque.com/mcp as a connector, and a Panelque page opens asking which store to connect. You sign in with Shopify, approve a screen that lists exactly what the assistant will be able to read, and that's the connection. Nothing to copy, nothing to keep safe.

We ask you to sign in with Shopify because it's the only way to be sure the person approving the connection is the one who runs the store — the same check that gates the app itself.

Either way the result lands in one list. Make one per assistant, and when you lose a laptop or stop using something, revoke that one and the others keep working.

The tool that earns its place

Every other tool answers the question you asked. One answers the question you didn't.

check_data_health looks at the same period and reports what's wrong with your own records – specifically, the things that don't produce a wrong-looking number but a flattering one:

• Most orders carry a packing and postage cost, but eighty of them don't. You aren't shipping eighty orders for free; the rows are missing, and the margin on all eighty is overstated.
• An expense category that recorded something every month and then recorded nothing this month. If that cost still happened, this month's profit is overstated by exactly the amount nobody entered.
• Products with no cost on their recipe. Their margins look wonderful, and so does every total built from them.

It reads the most recent quarter of the period separately, so a process that broke last week reads as a break in progress rather than being diluted by the thirty days before it. Each finding says which direction it bends your numbers, so the assistant can tell you a figure is unreliable instead of repeating it confidently.

Ask for the cost analysis first. Then ask it to check data health for the same period before you act on anything.

The rules we settled on

• Read-only, permanently. An assistant can look at everything and change nothing – not in Panelque, not in your Shopify store.
• The numbers match the app. Same currency, same VAT basis, and orders beyond your plan's daily limit are left out here exactly as they are everywhere else.
• A token is a password. Anyone holding one can read this store's data. Revoking takes effect immediately, whether it was pasted or signed in.
• Signed-in connections renew themselves. They expire after thirty days and refresh in the background, so you're asked once and not again.
• Recent months flatter themselves. Refunds land weeks after the sale. The assistant will tell you this if you ask it to; we'd rather say it here too.

AI access is included with every paid plan.

Why we built it

We kept catching ourselves writing the same three questions into the app as new pages. A page is the right answer when everyone needs it every day. It's an expensive answer when one person needs it once, in March.

The honest version is that the assistant is often wrong about what a number means – it will happily explain a margin that's only high because a cost is missing. That's why check_data_health exists, and why it was the second thing we built rather than the last.