There is a particular kind of small task that dashboards are bad at. Not difficult tasks — small ones. The ones that take four clicks, come up unpredictably, and always arrive while you are in the middle of something else.
Re-pointing a QR code is exactly that shape. The autumn menu goes up, and somewhere there are forty table cards with a code on them pointing at the summer one. The change itself is trivial: the codes are dynamic, that is the entire reason they were used. But doing it means finding the login, finding the codes, editing each one, saving each one.
So we did something about the interface rather than the feature. You can now connect Claude, ChatGPT, or whatever assistant you already have open, directly to your account — and say it instead.
What this actually is
It is an MCP server. MCP — the Model Context Protocol — is an open standard for letting an AI assistant call a defined set of operations on a service, on your behalf. Anthropic published it, and it has been picked up well past Claude: ChatGPT supports it, so do Cursor, Windsurf, Zed, Cline and a growing list of editors.
The important word is defined. This is not the assistant logging into your account, driving a browser, or holding your password. It is a fixed list of operations we publish, and nothing outside that list exists as far as the connection is concerned. You paste a URL and a token into your assistant once, and it can do those things and only those things.
What you can ask for
The list is deliberately short, because the useful things here are few:
- Re-point a code. "Point the café menu code at this PDF." Every copy already printed follows within seconds. This is the one that earns the feature.
- Ask about scans. "Which of my codes got the most scans this month?" Or the better question, "which have never been scanned?" — a code with no scans is either somewhere nobody looks or printed wrong, and both are worth knowing.
- Create a code. A name, a destination, optionally a folder and a custom short link. It comes back with the URL ready to print.
- Pause and resume. The landing page is broken; take the code out of service, put it back afterwards.
The scan figures it quotes come from the same code that draws your dashboard, over the same periods, with the same bot filtering — link previews, crawlers and uptime monitors excluded from every count. An assistant quoting different numbers from your own screen would be worse than no assistant at all.
What it cannot do, and why that is structural
It cannot delete anything.
Not "we have asked it not to" — there is no delete operation published on the connection at all. An assistant can only call what it is offered, and removing a code is not on the list. This matters more than it might sound: the whole risk model of giving a language model access to a real account is that it might misread a sentence, and the difference between a misreading you can undo and one you cannot is the difference between an annoyance and an incident. Deleting stays in the dashboard, where there is a trash you can restore from for thirty days.
It also cannot see or change your password, reach your billing details, or touch any other account. Every operation runs against your own records, so somebody else's code is not "access denied" — it does not exist as far as your connection can tell, which is the only version of that answer that leaks nothing.
Each assistant gets its own key
You create a token per assistant and name it after where it lives: "Claude on my laptop", "ChatGPT". They are separate credentials, revocable separately, so unplugging a machine you no longer have does not unplug everything else.
The token is shown to you exactly once, when you create it. We store a one-way hash and nothing else, which means we genuinely cannot show it to you again — and it also means a copy of our database is not a set of live keys to customer accounts. If you lose one, make another and revoke the old one.
Your dashboard lists what has actually connected: which assistant, which version, when it was last active. That list is evidence rather than configuration — a row appears because something completed a handshake and named itself, not because anybody ticked a box.
Setting it up
In Claude it is a settings screen: Connectors, add a custom connector, paste the URL and the token. In editors it is a small block of JSON, which your dashboard prints out ready to copy with a placeholder where the token goes. Either way it is about two minutes, once per assistant.
There is a full setup guide covering each client, and a page explaining the feature in more detail.
Who it is for
Honestly: people whose codes are already printed. If your codes live on a screen you can edit, this saves you very little. If they are on menus, packaging, window vinyl, event signage or anything else that costs money to reprint, then the gap between "the destination needs to change" and "the destination has changed" is exactly where this helps, and it is a gap that is usually measured in days rather than minutes — not because the change is hard, but because nobody got round to it.
It is available on the Pro and Team plans.