> ## Documentation Index
> Fetch the complete documentation index at: https://docs.useagentbank.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Instructions for your agent

> A concise operating contract for agents using AgentBank MCP.

Use the canonical [AgentBank Pay skill](https://useagentbank.com/SKILL.md)
for the complete workflow. The following is a readable summary.

```text theme={null}
Before AgentBank work, inspect the available MCP surface and call whoami first.
If local onboarding tools are present and credentials are missing, begin
onboarding once, show the browser authorization URL, and wait with the same
enrollment ID. If hosted spending-grant tools are present, use OAuth identity
and never run local onboarding.

Use AgentBank MCP tools rather than direct HTTP requests or hand-built payment
payloads. Use canonical asset objects and decimal-string amounts. Never infer a
recipient, wallet, bank account, token contract, chain, decimals, or calldata.

Estimate without a recipient. Before durable creation, collect a recipient
that satisfies the returned recipient requirements and show the complete
recipient, source and destination amounts, every fee and fee currency, route,
and expiry. Follow the returned approval status; do not infer a fixed threshold.

Use stable request IDs for identical logical retries. Execute only the current
Core-owned instruction. A transaction receipt is not final payment completion;
trust get_payment. On hosted OAuth, keep approval, funding, and progress cards
separate, and require a new user choice before spending-grant funding. Never
expose partner identity, secrets, Privy tokens, wallet keys, or World ID proofs.
```

<Warning>
  This summary does not replace the canonical skill or the live MCP instructions.
</Warning>