Set up an agent to pay

An agent that pays needs three things: guidance on how x402 payments work, a tool that signs and pays, and spending limits that the tool enforces. Skills provide the guidance. radius-cli or radius-sdk performs the payment. Neither the model nor its instructions should hold the signing key.

Add Radius skills

For agents that support the skills installer, run:

npx skills add radiustechsystems/skills

Choose your agent in the installer and add the guidance its task needs:

• x402: request paid resources and handle payment challenges, signing, and settlement.
• dripping-faucet: get testnet funds.
• radius-dev: build an integration, when the agent also writes code.

If your agent does not use the installer, add the linked Markdown files through its instructions or context. See the skills repository for other installation options.

Skills do not install a payment tool or enforce a spending limit. Configure those separately, below.

Choose a payment tool

Both pay only in SBC on the configured network and refuse an offer above the limit before signing (the CLI with -y or without a terminal; otherwise it asks).

Create a project wallet

For the CLI, keep the wallet in the project and select the network explicitly:

export RADIUS_HOME=.radius
export RADIUS_NETWORK=testnet
radius-cli wallet address
radius-cli wallet balance --json

The first command creates the wallet. Keep .radius/ out of version control, and give the agent only the wallet access it needs. See project-scoped agent wallet for passwords and other options. For application code, pass the signer to createRadiusFetch from a secret store.

Fund the wallet with testnet SBC: see create and fund a wallet.

Set spending limits

Decide which endpoints the agent may pay, and the most it may pay per request.

• Per request: --x402-threshold with -y is a hard cap for the CLI; the CLI refuses a more expensive offer instead of prompting. maxPerRequest does the same in radius-sdk.
• In total: neither per-request limit caps spending across requests. In application code, enforce a total budget or a seller allowlist in onPaymentRequired; see limit what you pay. With the CLI, enforce it in the tool runner that calls the command.

Enforce limits in the component that signs. Account for concurrent requests, and keep the amount spent across restarts. Software with direct access to the key can bypass limits implemented in another client.

Connect the agent

Expose the CLI command or your payment function to the agent through its tool runner. Set the network, wallet, allowed endpoints, and limit in the tool, not in the prompt. Use --json so the agent can read the result, including the payment. Keep private keys and wallet passwords out of prompts and logs.

Before the first payment, have the agent check the balance and state the endpoint and price it intends to pay. Then follow pay from a terminal for a first paid request.

To have a coding assistant help you build an integration instead, see coding assistants.