Payment guardrails for AI agents

Let the agent work.
Keep your hand on the till.

Countersign sits between an AI assistant and your PayPal account. Every refund and payment the assistant attempts is checked by plain code, not by another model. It runs, waits for your approval, or is refused.

Open the live shopWatch a recorded day
CheckedEvery refund and payment the assistant attempts goes through the guard first.
ApprovedAnything unusual waits for you, with the reason, until you decide.
RecordedEvery attempt is written to a log that cannot be edited without it showing.
YoursYour limits, your suppliers, your PayPal account. Nothing is built in.
Working now

The live shop

A working shop with its own orders, suppliers and PayPal account. Send it an email and an AI assistant handles it. Every refund or payment it tries goes through Countersign, and everything below updates as it happens.

  1. 1. Buy somethingPlace a test order below with your own email address. It is a real payment in PayPal's sandbox, with no money involved.
  2. 2. Ask for a refundWrite to the shop from that address (use the message form below). The assistant checks the order, PayPal makes the refund and the reply comes back.
  3. 3. Try to cheatAsk for a refund on someone else's order, or tell the shop a supplier has new bank details. Watch what the guard does.
  4. 4. Ask for something bigA refund over the limit waits in the owner's queue, and the owner gets an email with a link to approve or refuse it.

Be a customer

Buy something from the shop with your own email address. It is a real payment in the PayPal sandbox, with no money involved. Then write to the shop from that address and ask for a refund.

Inbox

Write as a customer, a supplier or the owner. The assistant reads it and replies.

Waiting for you

The guard holds anything it will not run by itself. Approving re-runs the guard first.

Activity

Every refund and payment attempted, with the guard's verdict.

Real runs, replayed

A day at the shop

The same inbox through the same AI agent against the PayPal sandbox: once with no guard, once with Countersign. Step through the emails and watch where the money goes. A step marked failed was turned down by PayPal's sandbox itself, for example when the test account had run short of funds. It is not a decision of the guard.

Loading the recorded days...
Hands on

Try to fool the guard

Pretend to be a customer, a supplier or the owner. Choose who is writing, what they ask for and whether their email can be trusted, then see what the guard decides and why.

The verdict appears here, with the reasons the guard gives for it.

For your shop

Set it up in minutes

Connect your PayPal account, say what the assistant may do, tell it about your shop and list the suppliers you pay. Nothing about a particular shop lives in the code.

Built to be trusted

How your money stays safe

The assistant reads email from people you have never met. These are the rules the guard applies to everything it tries to do.

Six checks on every refund and payment

  • IdentityThe mail server's SPF, DKIM and DMARC result decides who really sent a message. Only a confirmed message from the owner's address is the owner. Anything that fails is held.
  • AuthorityA customer can only ask for a refund on their own order, and a supplier's invoice is paid only on the word of a contact on record for that supplier.
  • Need to knowLookups return nothing about orders, payments or invoices that do not belong to the sender, so the assistant cannot repeat what it never saw.
  • ProvenanceEvery amount and payee is traced to where it came from. A figure that appears only in text a stranger wrote does not move money.
  • Limits and approvalThe owner sets the limits. Anything above them waits for approval, and approval re-checks the guard before anything is sent.
  • AuditEvery attempt is written to a hash-chained log, and each reply is checked against it before it is sent.