smolagents tool: payment requests that only a human can sign

[send21io]

I'm with send21 (disclosure).

Short one, code first. A CodeAgent writes and runs Python, so whatever its tools can do, it will eventually try. For anything involving money I want the tool's maximum power to be "write a draft".

from smolagents import CodeAgent, tool

@tool
def payment_link(amount: float, fiat: str, ref: str) -> str:
    """Create a pay link the payer signs in their own wallet.

    Args:
        amount: amount in fiat, e.g. 25.0
        fiat: ISO currency code, e.g. "SEK"
        ref: your order or invoice id
    """
    # mcp_call(): tools/call on https://send21.io/mcp (helper, not shown)
    r = mcp_call("create_payment_request", {
        "fiatAmount": amount, "fiatCurrency": fiat,
        "options": [{"currency": "Btc", "network": "Mainnet",
                     "address": "me@wallet.com", "method": "Lightning"},
                    {"currency": "Sol", "network": "Solana", "address": MY_SOL}],
        "orderId": ref, "memo": None, "expiryHours": None, "idempotencyKey": ref,
    })
    return "https://send21.io" + r["payPath"]

agent = CodeAgent(tools=[payment_link], model=model)

Why it holds up even if the agent loops:

• The send21 key is scoped to drafts. It cannot move funds, and send21 never holds them.
• Same ref = same idempotency key, so a retry returns the original link instead of a second one.
• Lightning to your own address confirms automatically when your wallet supports LUD-21; on-chain is confirmed by an HMAC-signed webhook.

Not an official smolagents integration, just @tool.

Question: do you sandbox CodeAgent tools with side effects differently (separate executor, approval step), or rely on the tool itself being harmless?

Try it on testnet, no account: https://send21.io/demo

send21, info@send21.io