Skip to content

Learn

How an agent pays. Your money, your rules, its errands.

An AI agent can pay for a service without your card and without an account at the provider. You fund a wallet, you set the rules, and the agent pays per call inside them.

Updated

The problem.

Ask an agent to make a product page and it will want a logo, a few images and a screenshot of a competitor. Each of those is a paid API at a different company, and each one wants an account, a plan and an API key before it does anything.

Handing an agent your card or a pile of API keys solves that badly: there is no limit on what it spends and no record of why. Agent payments fix it with three parts: a wallet the agent can spend from, rules that bound the spending, and a way to find and pay services in the middle of a task.

How it works in Sourcy.

  1. 1A person funds a wallet. You create a wallet for your agent and add USDC. When it runs low, the agent asks you for more with a QR code.
  2. 2A person sets the rules. A maximum per purchase, a daily cap and the categories the agent may buy from.
  3. 3The agent connects once. One MCP URL in Claude Code, Claude, Cursor or another MCP client gives it tools to search services, buy one and check its wallet.
  4. 4It finds a service and sees the price. It searches the catalogue by what it needs, for example a screenshot, and compares providers and prices per call.
  5. 5Sourcy checks the rules, then pays. If the purchase breaks a rule it is blocked and recorded with the reason. Otherwise Sourcy pays the provider's wallet with x402 and returns the result with a receipt.

What the agent sees.

To the agent, paying is a tool call. It asks for a service with a price ceiling and gets back either the result and a receipt, or a plain reason why it was not bought.

Example tool call, shortened
buy_service({
  slug: "<provider>",
  operation: "capture",
  input: { url: "https://example.com" },
  max_price_usd: 0.01
})
→ result + receipt { price_usd, tx_hash, pay_to, latency }

What stays with you.

The rules are checked before any payment, not after. Every purchase, including blocked ones, is recorded with the provider, the price and the transaction, so you can see what the agent bought and why.

If the provider's call fails, the payment is not settled and the agent is not charged. Payments are in USDC on Base, which is real money.

Questions. Short answers.

Does the agent get my card or my API keys?

No. It spends from its own wallet, inside the rules you set. It never sees a card or a provider's API key.

Can the agent overspend?

Not past your rules. A purchase over the maximum per purchase, over the daily cap or outside the allowed categories is blocked before any payment and recorded with the reason.

Which agents can use it?

Any agent that speaks MCP, such as Claude Code, Claude and Cursor. Agents without MCP can call the same HTTP API.