Founder or CTO of an AI-agent company
Uses Onbehalf to get deals past security, risk and finance review. This is who we sell to.
Onbehalf sits between your AI agent and your money. Your rules decide. A person signs off when needed. Onbehalf is the notary: it seals a receipt anyone can check.
The person who owns an agent writes its mandate: which actions it can take, up to what amount, and for payments, who it can pay. Every agent has an owner and a backup owner.
Every request gets one of four answers: yes, yes but flag it, ask a person, or no. No AI makes the call. In our test workspace, an agent can pay up to 1.00 USDC on its own, up to 3.00 with a flag for review, and up to 10.00 if a person says yes. Above that, the answer is no. On Base Sepolia, a person says yes to every payment. Paying someone who isn't on the list is always a no, even for a small amount.
The approver sees the amount, the reason and why it needs them. Their yes is locked to that exact action. If anything changes, the yes no longer counts. After a yes, there are 10 seconds to undo. On Base Sepolia, the yes is a signature from the approver's wallet. The payment is sent from that same wallet. On Stripe, the yes needs the approver's passkey.
The payment goes out once, never twice. For Stripe, Onbehalf sends it. On Base Sepolia, the approver's wallet sends it, and Onbehalf watches the chain until it is final. If the rail's answer is unclear, Onbehalf checks the outside record before it tries anything again.
Onbehalf signs a receipt. It names the mandate and holds the rules' decision, the approval and the outcome. Change one detail and the check fails.
Anyone with authority over an agent can stop it, instantly. Stopped agents can't move money until someone with the right to do so turns them back on. The stop itself is signed, so it can be checked later too.
What was allowed, who said yes, and what happened. Check it on the Verify page.
The payment's own record at the rail. For Stripe, the Stripe record. For a chain, the chain record: two public explorers, linked from the receipt's own chain id and transaction id.
Proof of when the receipt existed. Planned.
Uses Onbehalf to get deals past security, risk and finance review. This is who we sell to.
Connects the agent. The agent can ask for a payment but can never approve one.
Sees each agent's limits, approves payments and can stop any agent, all from one Console.
Checks receipts on the Verify page, without trusting us.
Built and working.
Built. The approver's own wallet signs and sends the exact payment. Every payment here needs a person. First one sent Oct 8, 2026.
Refunds, charging a customer's saved card, and account credits. Test money only, and the money stays in your Stripe account. Paying suppliers runs on USDC today; bank payments to suppliers are next.
Built and proven with a real Claude agent. Invited design partners connect their own agent. Opening it to everyone is next.
Built. Download every signed receipt as a .zip, with a CSV list. Invite your auditor to a read-only seat. A designed PDF book is next.
Planned.
Built. Invited design partners get their own workspace at app.getonbehalf.com and connect their own agent.
Built. A one-time code by email. Approving a payment needs a passkey.
Fixed rules and people decide. AI agents only ask.
Receipts are signed, and anyone can check them without us.
Onbehalf doesn't hold your money. Payments run through your own accounts, on the rails you already use.