Software that spends money needs a line it can't cross.

For thirty years, spending money online meant a person was present. A card was typed in, a button was clicked, someone was accountable for the click. Agents break that assumption. They decide at runtime what to call and how often, and there is nobody at the keyboard when they do it.

The usual answer is to hand the agent an API key and put the company card behind it. That works until it doesn’t: one limit covers everything, there is no way to stop one agent without stopping them all, and a key sitting in an environment variable is one prompt injection away from being someone else’s.

What we build

A wallet per agent, with a daily cap, a per-transaction cap, and a threshold above which a human signs. Keys are split 2-of-3 across enclaves and never assembled where your code runs. Payments settle in USDC and USDT across six chains, and gas on Base is on us.

You drive it from the command line. That is the whole product; there is no second thing we are quietly also selling.

What we won’t do

• Hold a whole key anywhere, including on our own machines.
• Let a payment through after the cap is spent.
• Charge a seat fee, a minimum, or a setup fee.

How we work

We’re small on purpose. A payments product is judged on what it refuses to do, and fewer people means fewer things ship that nobody can defend.

Every feature gets designed against one question: what happens when this is wrong at three in the morning with no human watching? That’s also why we can’t sign on your behalf — we never hold a whole key. It’s an architecture decision, not a policy we could quietly change later.

Where we are, and hiring

Early access is going out in batches while we work with the first teams putting agents on real budgets. If that’s you, the fastest way in is the docs.

We’re hiring engineers who have shipped systems that move money and can argue about failure modes without getting bored. Write to hello@agentwallex.com with something you built and what broke.