In plain words

What is this, actually?

No jargon on this page. If you have ever worried about what a computer program might do while nobody is watching it, this explains what we do about that.

Start with the thing that goes wrong

Imagine you ask an assistant to refund a customer $240.

The assistant sends the refund. Then the internet hiccups, and the reply never comes back. The assistant is now stuck on a question it genuinely cannot answer:

"Did that refund go through, or not?"

It has two choices, and both of them are wrong:

Try again

If the first refund did work, the customer now gets $480. Your money, gone twice.

Give up

If the first refund didn't work, the customer gets nothing and complains tomorrow.

This is not the assistant being stupid. A person would be stuck too. The information simply isn't there.

What we do about it

We keep the notebook that the assistant cannot keep for itself.

Before it does anything it cannot undo, the assistant asks us one question: "Am I allowed to do this?" We write the answer down somewhere permanent before it acts — so even if the assistant crashes, restarts, or forgets everything, the note survives.

The assistant asks Ratchet before acting. The second attempt is refused as already done. FIRST TRY Assistant "may I refund?" Ratchet writes it down yes, go Refund sent $240 IT PANICS AND TRIES AGAIN Assistant "may I refund?" Ratchet checks the note already done Nothing happens customer keeps $240
The second attempt is stopped by the note, not by luck.

That's it. That's the whole product. We never touch your money, your accounts, or your customers — we only answer the question. Your own software still does the work.

Setting it up, in three steps

You need someone comfortable editing your software's settings for step 3. Steps 1 and 2 take about a minute and need nothing technical at all.

  1. Make an account

    Go to the sign-up page and enter a name and your email. You get a key straight away — a long password-like string starting with rk_. It appears once. Copy it somewhere safe, like a password manager.

    No card. Nothing to cancel.

  2. Click the link in your email

    We send one message. Clicking it confirms the address is real and switches on the free allowance — 1,000 protected actions every month.

    Your key already works before you click. Confirming just raises the ceiling.

  3. Give the key to whoever built your AI assistant

    That is genuinely the whole handover. Send them this page and the key:

    Setup instructions for developers →
    Works with Claude, Cursor, ChatGPT tools, n8n, LangChain and anything that can make a web request.

What it protects you from

Anything your assistant does that you cannot take back. For example:

  • Charging or refunding a customer twice because a payment page timed out.
  • Sending the same email to your whole list twice because the first send looked like it failed.
  • Publishing or deploying the same thing twice, which can take a site down.
  • Spending more than you meant to — you set a limit per job, and the assistant cannot raise it.
  • Not knowing what happened overnight — everything it did is written down and readable.

What it does not do

We don't decide whether something was a good idea — only whether it has already happened. If your assistant decides to refund the wrong customer, we will faithfully make sure it does that exactly once.

And when an action's outcome is genuinely unknowable, we say "unknown" rather than guessing. A guess that sounds confident is worse than an honest shrug, because you would act on it.

Still not sure?

The questions page answers most things in plain words too. Or just make a free account — it costs nothing and there is nothing to cancel.

Create a free account Read the questions