The Infrastructure Layer for Autonomous AI Agents

Give your AI agents real-world financial capabilities. Safely.

Non-custodial, policy-driven middleware connecting LLM frameworks to real bank accounts via Plaid and Lithic. Zero pre-funded wallets required.

connect codex · claude · grok · gemini → your bank

View documentation
agentbank · runtime tracelive
01agent_99 › tool_call issue_card { max_usd: 25.00 }
02

the gap

Built for autonomous loops, not human popups.

consumer payment tools

  • Pauses the loop waiting on a human tap-to-approve
  • Pre-funded wallets that idle capital between runs
  • Checkout flows built for browsers, not background workers
  • Receipts with no link to the prompt that caused them

agent bank

  • Fully programmatic: every action is a tool call
  • Real-time JIT authorization decisions in ~200ms
  • Policy engine evaluates before funds ever move
  • Prompt-correlated telemetry on every authorization

execution model

From prompt to payment in milliseconds.

How autonomous workers seamlessly interact with your treasury infrastructure without breaking loop execution.

01

Tool-first integration (MCP server)

Agents natively discover and call financial tools straight out of the box using standard MCP or SDK imports in Cursor, LangChain, or custom loops.

issue_card()check_budget()pay_link()
02

Zero human bottlenecks

Built for background workers, data scrapers, and multi-hour cloud operations. No pinging a phone to approve a $12 API fee — the agent operates inside strict, developer-defined daily budgets and policy guardrails.

daily_cap()mcc_lock()auto_pause()
03

Settle, trace, replenish

Every authorization is written back with its prompt trace, then operating liquidity is topped up in the background so the next run never blocks.

trace_id()ach_debit()acss_pad()

core architecture

Three moving parts. No custody.

Funds stay in your bank account until an agent authorization clears policy.

plaid

01One-click bank auth

Instant OAuth linking for US & Canadian checking accounts with live balance validation before a single card is issued.

lithic

02Instant JIT card generation

Single-use virtual cards created on-the-fly, backed by real-time authorization webhooks in roughly 200 milliseconds.

ach / pads

03Automated rebalancing

Background debits via Stripe ACH (US) and ACSS PADs (Canada) to replenish operating liquidity between agent runs.

developer experience

Mount it in one import.

Same policy engine behind every surface — SDK, REST, or MCP.

import { AgentTreasury } from '@agent-bank/sdk';

const client = new AgentTreasury({ apiKey: process.env.AGENT_BANK_KEY });

const card = await client.issuing.createVirtualCard({
  agentId: 'agent_99',
  maxBudgetUSD: 25.00,
  promptTraceId: 'task_abc123',
  mccLock: ['5734'],
});

// -> { last4: '4417', state: 'OPEN', expires_in: 900 }

capabilities

Treasury primitives for agent fleets.

Prompt-correlated telemetry

Trace every dollar spent directly back to the exact LLM prompt execution step — agent, run, tool call, and token cost on the same ledger row.

task_abc123#4issue_card$12.00
task_abc123#9pay_link$3.40
task_zf910#1issue_card$25.00

Cross-border North America

Native support for USD via ACH and CAD via Rule H1 PADs, with per-currency policy ceilings.

USD · ACHCAD · PAD

Granular guardrails

Hard daily caps, merchant category code locks, and auto-pause triggers — evaluated server-side on every authorization request.

daily_cap_usd

200.00

mcc_allowlist

5734, 7372

auto_pause_on

3 declines

pricing

Priced per platform, not per prompt.

Sandbox

$0

test keys, simulated rails

  • Unlimited test cards
  • Full MCP tool surface
  • Community support

Scale

$99

per month + interchange rebate

  • Live JIT issuing
  • ACH + PAD rebalancing
  • Prompt-trace exports
  • Slack support

Enterprise

Custom

dedicated policy cluster

  • Private BIN range
  • SOC 2 evidence pack
  • 99.99% authorization SLA