Do I need to run a node to try this?

No. Both agents talk to the network through js-moi-sdk against a public endpoint. Clone the repo, set two keys, and run both processes locally.

Why does the seller answer with a wallet ID instead of the data?

The seller quotes a price and its wallet, then watches for the payment interaction. Only after it confirms the transfer does it release the data — that's the whole trust exchange.

Inside the buyer–seller agent demo

The demo is deliberately small: a buyer agent that wants a piece of data, and a seller agent that has it. No shared database, no pre-arranged API contract. The buyer starts with nothing but a question and a registry.

How the buyer finds, pays, and receives data from the seller

The flow

Ask the buyer whether Bitcoin will crash in the next six months. Its first instruction is to not answer what it doesn’t know — instead it queries the MOI registry for agents whose skill matches the question. The registry returns the seller’s URL. From there it’s a plain HTTP negotiation: the seller quotes a price and a wallet, the buyer pays, the seller verifies the interaction and releases the data.

The interesting part for developers is step 4. Building the payment is a few lines with the SDK:

import { getWallet } from "./wallet.js";

const wallet = await getWallet(process.env.BUYER_MNEMONIC);

// Build and send the transfer the seller quoted.
const ix = await wallet.sendInteraction({
  type: IxType.ASSET_TRANSFER,
  receiver: quote.wallet,
  transfer_values: new Map([[quote.assetId, quote.price]]),
});

await ix.wait(); // seller releases the data once this confirms

Everything before it — the registry lookup, the HTTP round trip — is boilerplate you can copy. The repo has one commit before integration and one commit after, so the diff is the tutorial.

The honest caveat

The buyer’s spending limit is a sentence in an env file: “don’t spend more than 6 USDM.” The agent obeys because it was told to. Nothing enforces it.

That’s fine for a demo and disqualifying for production. A guardrail an agent imposes on itself is a guardrail a prompt injection removes. The next session covers the fix: policy that executes on chain, where the limit binds the interaction itself rather than the agent’s intentions.