AgentBank is where businesses open accounts built for AI agents to hold and spend money on their own. Every agent is identified, every dollar has a limit, and you can freeze any of them in one click.
No screen-scraping a human's login. No sharing one corporate card with ten bots.
Your business gets a real account with a routing and account number. Each agent gets its own account underneath it, so its money is never mixed with anyone else's.
Each agent is registered with a verified email inbox, the IP address it runs from, what it is for, and what it may spend. Its key works only from that IP, and it cannot pay until its inbox is verified.
Give an agent a virtual card that draws only on its own account. If the agent is compromised, the damage stops at its balance and its limits.
Agents pay by API or by card. Each payment is checked against the agent's limits before any money moves, and declines are recorded with a reason.
The agent holds a key that can only spend from its own account, within the limits you set. The rules live at the bank, so a prompt injection cannot talk its way past them.
# the agent pays a vendor with its card curl https://api.agentbank.example/v1/payments \ -H "Authorization: Bearer ab_test_…" \ -H "Idempotency-Key: order-1042" \ -H "Content-Type: application/json" \ -d '{ "amount": 31200, "card_id": "card_2x8g…", "merchant": "CLOUD COMPUTE CO" }' 201 { "status": "completed" } # the same agent tries to overspend 403 { "status": "declined", "reason": "daily_limit_exceeded" }
Traditional banks treat an agent logging in as fraud, because they cannot tell an authorized agent from an attacker. AgentBank starts from the other end: the agent is a known, limited, accountable operator on the account.