Anchor Browser adds x402 payments for agents via AgentCore

AWS published a post on 13 October 2026 describing how Anchor Browser, a cloud browser for AI agents, pays for web resources that answer with HTTP 402 Payment Required. The integration uses Amazon Bedrock AgentCore payments and the open x402 protocol. The post was written by AWS and Anchor staff and is vendor-published. It reports no customer results.

How the payment works

When a browser session starts with x402 turned on, Anchor Browser creates an AgentCore payment session tied to a payment instrument and a budget. If the agent hits a 402 response, the browser matches the merchant's payment method to that instrument. AgentCore checks the amount against the session limit and signs the proof through a wallet provider, either Coinbase Developer Platform or Stripe Privy. The browser then retries the original request once with the signed header.

The post says no key material enters the browser or the agent runtime. Playwright, Puppeteer and Chrome DevTools Protocol scripts need no payment code, and the page loads as if it were free. Payment attempts, approvals, denials and settlements are logged to Amazon CloudWatch, according to AWS.

What a team controls

• One switch: session.x402 defaults to false, and the post says setting it to true authorizes real payments.
• A budget and expiry per browser session, enforced by AgentCore before it signs. If the budget is exhausted or AgentCore refuses, the page receives the original 402 unchanged.
• The scope: the flag pays third-party resources only. It does not pay for the browser session itself.
• Billing: Anchor settles third-party x402 charges in credits on the same invoice as browser time, so finance does not handle a stablecoin balance, per the post.

What the post leaves out

• The limits. The post says the specific spending and time limits are in Anchor's technical documentation and gives no figure.
• Any price or fee for the Anchor integration, or any markup on the credits.
• Any failure data: how often payments are denied, retried or settled for the wrong amount.
• Prompt injection. The post lists it as a risk of uncapped spend and relies on the infrastructure-level limit, but gives no test of it.

The practical reading: the limit, not the model, is what stops a looping or manipulated agent from spending. If your agents browse paid sources, set a small per-session budget first, keep the flag off for workflows that do not need it, and check the CloudWatch trail after the first week. We covered the same AgentCore payments layer in a BlockRun inference-payments story earlier this month.