Observe. Control. Analyze.

The control room
for every AI agent
that moves your money.

ctrl payments sits between any agent and the money it moves. Every attempt is on the record, your rules decide what goes through, and you see the pattern before anyone else does. Built on ctrlrun, open source.

Sign up Discuss your deployment

Control. Every payment your agents attempt, seen live, before it moves. Your rules run before the payment does. One yes. Exactly that amount. Agents can retry. Your refund shouldn't. The morning's receipts, refused ones included, become evidence you can hand to finance. Ask in plain language and get answers backed by the receipts. You see the pattern before anyone else does. Sample data

$312.50 support-agent refunds a customer seen, 09:58allowed, under the line

$4,000.00 support-agent proposes a refund to a customer it has never refunded seen, 10:12 held for finance-approver committed p.rao approved exactly $4,000.00. Not a cent more. Slack, 10:14

$5,000.00 support-agent executes the same refund under that approval refused, mismatch

Receiptrcpt_01J9KX

agentsupport-agent, through the gateway

paymentrefund $4,000.00 USD on Stripe, ch_1Pq

decisionheld, then approved by p.rao, policy v14

outcomeCOMMITTED, reconciled with the rail

chainprev 3f9c1e…a1 this 8be27d…7d

verifychain intacton your machine, without us

Evidence, this morning

r_9f0allowed, under the line

apr_31fapproved exactly $4,000.00

r_9f4refused, mismatch

chain intact, verified 16 of 16

Agent behavioursupport-agent, this week

you askWhat did support-agent refund this week, and to whom?

10:02$480.00allowed, under the $500 line

10:21$450.00allowed, under the line again

10:40$470.00refused, the seven-day budget for this customer is spent

splittingowner notifiedthree refunds under the line to one customer in forty minutes, named as a pattern

answered from 41 receipts Refunds to cus_4Ab, a day, this week.

Know exactly what your AI agents are paying,
everywhere you gave them access.

Observe.

Every payment your agents attempt, seen live, before it moves.

Every attempt, refused ones included, leaves a receipt: tamper-evident, across every rail, verifiable without us.

Payments feedSample data

09:58support-agent refunds $312.50allowed, under the line
10:12support-agent refunds $4,000.00 to cus_9Kx, never refunded beforeheld for finance-approver
10:14p.rao approved exactly $4,000.00committed
10:15the same refund at $5,000.00, under that approvalrefused, mismatch

Receiptrcpt_01J9KX

agent: support-agent, through the gateway
payment: refund $4,000.00 USD on Stripe, ch_1Pq
decision: held, then approved by p.rao, policy v14
outcome: COMMITTED, reconciled with the rail
chain: prev 3f9c1e…a1 this 8be27d…7d
verify: chain intacton your machine, without us

Control.

Your rules run before the payment does. Inside the band it goes through. Above it, a person decides. Outside it, nothing leaves.

Budgets per agent, per counterparty, per day. A yes is for one exact amount. Payouts never go alone.

Payment policySample data

Can it pay a vendor on its own?

support-agentrefunds $312.50 to cus_9Kxallowed, under the line
support-agentproposes a $4,000.00 refund to a customer it has never refundedheld for finance-approver
invoice-agentproposes a $18,400.00 payout to a vendor, on its ownNo. Held for 2 of 3
procurement-agenttries a $2,400.00 purchase with no rule for itrefused, nothing leaves

One yes. Exactly that amount.

The finance lead approves exactly $4,000.00 from the tool they already use: the dashboard, their own assistant, or Slack. The approval is tied to that payment and that amount. Anything else needs a new yes.

Analyze.

What each agent tried to pay, to whom, how often, and where that is drifting.

Ask in plain language and get answers backed by the receipts.

Agent behavioursupport-agentSample data

Refunds to cus_4Ab, a day, this week. Today, three in forty minutes.

10:02refund $480.00, under the $500 lineallowed
10:21refund $450.00, under the line againallowed
10:40refund $470.00, the seven-day budget for this customer is spentrefused
signalsplitting: three refunds under the line to one customer in forty minutes, named as a patternowner notified
you askWhat did support-agent refund this week, and to whom?answered from 41 receipts

Whatever AI assistant you use for payments, just add this. Every payment is observed, controlled and analyzed, all in one dashboard.

The agent's MCP configuration changes one URL and loses one key. Free in test mode, no card.

ctrlpayments serve --public
# then add the URL it prints in the vendor's console,
# as a tool of your own
https://pay.your-company.com/mcp

pip install ctrlpayments

from ctrlpayments import Payments
payments = Payments.from_file("ctrlpayments.yaml")

@payments.protect("refund")          # one line above the function that pays
def issue_refund(customer, amount):
    ...

pip install ctrlpayments
ctrlpayments init --rail stripe
ctrlpayments serve
# add the URL it prints to your assistant's MCP config
{ "ctrlpayments": { "url": "http://127.0.0.1:8401/mcp" } }

All three land in the same dashboard. Every agent, every connector, every rail, observed, controlled and analyzed in one place.

Works with any integration, and we add integrations on request. Built on ctrlrun, the open-source kernel.

See pricing See use cases

Transparent by design

The kernel that decides and refuses is ctrlrun, open source under Apache-2.0. The receipt format and the payment model are published. Run the verification yourself, on your machine, against our product.

Trust is something you verify, not something we ask for.

Read the code on GitHub Docs at ctrlrun.dev

We do not move, hold, or sign money.We do not score your customers for fraud. We watch the agent, not the cardholder.We do not run KYC or sanctions checks. A connector may call a provider that does, under your policy.

No finance agent yet, or not sure what it should do?

On Enterprise we design and build it with you, controlled from the first payment.