Your AI agent can move money in Stripe - here's how to stay in control

In this Stripe developer blog post, Daniel Wood, a product manager on Stripe's Identity and Agents work, starts from what agents can now do through Stripe's MCP server, the interface that exposes Stripe actions as tools an AI agent can call: build checkout pages, run reports, issue refunds and adjust prices, and, using Stripe Treasury, send money to others. His answer to how a business stays in control is that the safeguard depends on whose identity the agent uses.

If the agent acts on your behalf, you connect it to the MCP server by signing in, then choose which accounts or sandboxes it can reach and whether it gets read or write access to each kind of resource. When that agent makes a refund or payout in live mode, Stripe requires you to confirm the action before money leaves the account, after which you tell the agent to try again. Wood says this confirmation applies only to certain write actions, but it holds even if you have switched off your agent client's own prompts asking permission before each tool call. In other words, the check sits on Stripe's side, not in whatever agent software you happen to run.

That sign-in route does not suit an agent that runs on its own, such as a bot that watches a support queue and handles refund requests, because the session is tied to one person's login and every confirmation goes to them. For that case Stripe offers agent-tagged API keys: long-lived keys that, like ordinary keys, support permissions, rotation and expiry, but are subject to approval rules. A rule names a trigger such as a refund being created, an optional condition such as an amount over $500, the reviewers, and whether to require approval or block the action outright. Stripe ships default rules for agent-tagged keys covering refunds, payouts and outbound transfers, though the post does not say what thresholds they use, and by default no one, human or agent, can approve their own request. When a rule applies, the API returns an approval_required error instead of completing the action and files a request that the agent can annotate with a reason. If a reviewer approves, Stripe itself attempts the action; unapproved requests expire after 14 days, and webhooks tell the agent how things turned out.