How it works
From an empty account to an agent that pays for what it needs.
Five steps. None of them hands anyone your keys.
Autonomy ends at the point of payment.
Agents plan, browse and call APIs on their own, right up to the moment something costs money.
Agents hit a paywall and stop
A per-use fee, an API that needs an account, a checkout. The workflow stalls until someone steps in.
Cards and API keys were not built for agents
Handing them to an agent means unbounded authority and no way to say what it may buy.
Nobody can tell which agent bought what
Finance sees a provider total, attributed to agents in a spreadsheet after the money is gone.
1 · Your account
An account only you control.
One Face ID prompt creates a Haven account you own outright. No seed phrase, no credit card. Add USDC when your agents need it, and a backup passkey when you like.
2 · Budgets
A budget per agent, not a card for all of them.
An amount and a period per agent, optionally pinned to one recipient. You sign it once, it refills itself, and you can stop it at any time.
3 · Connecting an agent
One command wires any agent in.
Paste the setup prompt to Claude, Codex, Cursor or any other agent harness. The agent's key is made on its own machine, Haven never sees it, and you approve the budget with your passkey.
I have a Haven account and I am signed in. Please set up Haven so you can pay for things within a budget I approve.
Start by reading /for-agents.md — it is written for you and explains which steps are mine.
Then run:
npx -y @haven_ai/connect@<channel> \
--setup hv_setup_… \
--api <api-url> --ack-local-tools- The agent requests a paid resourceGET api.research.example/query
- The merchant asks for payment402 Payment Required · 0.05 USDC on Base
- Haven checks it against the agent's budget214.00 used of 250.00 · recipient allowed
- The agent signs on its own machinethe key never leaves it
- The agent retries with the signed paymentthe merchant takes it from here
- Settled from your account to the merchantbudget enforced on-chain · over budget reverts, nothing moves
- The agent gets the resource200 OK · research.json
4 · When it pays
Seconds from paywall to result. No human in the loop.
Haven checks the price against the budget before the agent signs. The agent signs locally, and the chain enforces the budget: an over-budget payment reverts, and nothing moves.
5 · The receipt and the books
Every payment explains itself.
Every payment carries a receipt: which agent, what it bought, which budget allowed it, the on-chain proof. Connect your bookkeeping tool and it appears there as a draft invoice with the evidence attached.
Security
Built so that Haven cannot be the weak point.
Haven is non-custodial software, not a custodian or a payment processor. Even a fully compromised Haven cannot move your money.
Funds stay in your account
Haven never holds funds, or a key that could move them.
Limits are enforced on-chain
An over-budget payment reverts at execution. It is not a rule in our database.
A leaked credential stays capped
An agent's credential cannot spend your account past its budget. If it is exposed, rotate it: the agent keeps its history and the old credential stops working. A balance already in the agent's own wallet is controlled by its key, and rotating does not move it.
An exit that needs no Haven
Inspect and revoke every budget with only your wallet and a public RPC. Open the exit page
Recovery you control
A backup passkey replaces a lost one. Every signer change is signed by you, never by Haven.
Read the model
The full security model is public. Security model
Protocols
x402 today. MPP next. One budget either way.
Your budget, receipt and books do not change when the rail underneath does.
x402
Pay-per-request HTTP
A server answers 402, the agent pays, the request is retried with proof. Live.
Stripe MPP
Agent-initiated checkout
Subscriptions and purchases between an agent and a merchant. Next.
Set up your first agent.
One passkey, one budget, one command.