Stripe Treasury for agents lets ChatGPT and Claude move business money with human approval

What Stripe launched

On 21 October 2026 Stripe announced Stripe Treasury for agents, a way for businesses to run parts of their banking through AI agents such as ChatGPT and Claude. Stripe Treasury is the business account product that lets companies hold and move money from the same Stripe dashboard where they accept payments. The new layer lets an agent work with that account from inside the AI tools a founder already uses, instead of a person switching between a chat window and a banking site.

According to Stripe, agents can check balances, pay invoices, transfer funds between accounts and convert currencies. The company says agents will soon be able to do more, including creating cards, and that businesses will be able to set spending limits for their agents "in the coming weeks". The feature is labelled a public preview in Stripe's documentation.

The connection runs through Stripe's Model Context Protocol (MCP) server, the standard interface that lets an AI client call outside tools. Stripe's press release adds that it is not a bank and that checking accounts are provided by Fifth Third Bank, National Association, Member FDIC.

What an agent can actually do

The Treasury for agents documentation lists five Treasury-specific MCP tools, all in public preview, on top of Stripe's existing MCP tools. get_balance_summary shows a balance summary across the Stripe balance and Treasury accounts. convert_currency converts funds between currencies inside a Treasury account. send_money sends money from a Treasury account to a recipient, and setup_recipient creates that recipient. transfer_money moves money between supported Stripe balances.

Each tool maps to a named permission. Reading balances needs read access to balances and financial accounts. Sending money needs write access to outbound payments, setting up recipients needs write access to recipient configuration, and transfers need write access to outbound transfers and payouts. When a business connects an agent, it decides which of those capabilities to grant, so an agent can be limited to reading numbers and never be given the ability to pay anyone.

Setup is a one-line server address, mcp.stripe.com, added to an MCP client. The docs include install steps for Cursor, Claude Code, Claude Desktop, Codex, ChatGPT and VS Code. Readers new to the protocol can see how MCP tools and resources work in our explainer.

How the human approval works

The central design choice is that, by default, a person must review and approve any agent action that results in live money movement. Stripe's docs state this plainly for Treasury, and a 21 October post on the Stripe developer blog explains how it works in practice.

When someone connects an agent with OAuth, the agent acts on behalf of that person's Stripe login. The developer blog post by Daniel Wood, a Stripe product manager for identity and agents, says that if such an agent makes a refund or payout in live mode, Stripe requires the user to confirm the action before money leaves the account; after confirming, the user tells the agent to try again so it can finish. The post says this confirmation is required for certain write actions even if the user has turned off the AI client's own prompts asking for permission before tool calls.

That last point matters. Many coding and agent tools let users switch off per-step confirmations to work faster. Stripe's check sits on its own side, so a relaxed setting in the client does not let an agent move money silently. Administrators can also turn MCP access on or off for a whole team, separately for live mode and test sandboxes, and revoke any connected session from the dashboard.

Agents with their own keys and approval rules

OAuth is tied to one person's login, so it does not suit an agent that runs on its own, such as a bot watching a support queue. For that case Stripe offers agent-tagged API keys. They behave like normal API keys, with permissions, rotation and expiry, but are treated as an independent actor and are subject to approval rules.

Stripe's two-party approvals documentation, also a public preview, lets administrators require a designated reviewer to approve sensitive actions. Rules can trigger on actions such as a refund, a payment intent, a new subscription or a bank account change. Some actions can only be gated when an agent-tagged key starts them: currency conversions, outbound payments, outbound transfers and payouts. Conditions can depend on amount, on who is acting, or on how many requests arrive in a period, and the control can be either require approval or block.

Stripe also maintains default approval rules for agent-tagged keys, covering refunds, subscription cancellations, outbound payments, outbound transfers, payouts and currency conversions. When an agent hits a rule, the API returns an approval_required error and files a request automatically; the agent can attach a written justification. A user cannot approve their own request, and requests not approved within 14 days expire. Agents can listen for webhook events to learn whether a request was approved, rejected, expired, succeeded or failed.

A one-person company as the pitch

Stripe frames the launch around very small companies. Its announcement says more than twice as many solo founders earned over $1 million in 2025 as in 2023, and that 63 percent of C corporations formed through Stripe Atlas in the second quarter of 2026 had a single founder. Those are Stripe's own figures.

The example customer is Kubrio, a platform for children aged 6 to 13 to make films, games, comics and robot designs. Its founder, Vlad Stan, is the only employee. According to Stripe, his Grok-based agent, Stash, connects to Kubrio's Treasury account through the MCP server, checks balances, new and cancelled subscriptions, failed payments and refunds every morning, and sends him a report. Stan plans to let it transfer money between accounts and pay contractors and creators, with every payment approved by him. Crowdfund Insider carried the same details in its 24 October report.

"Every founder wants to spend more time building and less time on admin," said JR Farr, product lead for Stripe Treasury, in the release.

How it compares with other agent banking moves

Stripe is not first. Days earlier, the business bank Mercury said ChatGPT and Claude can request payments and transfers through MCP write actions, also with money movement approved inside Mercury. The pattern is the same: the agent drafts the action, and the account holder signs off in the bank's own interface.

Large consumer banks are moving the other way. Wells Fargo's new online banking terms treat whatever an agent does with a customer's login as the customer's own instruction while banks block outside bots, and American Express limits agents to its API and MCP. Stripe's approach sits closer to the second model: agents get a sanctioned door with scoped permissions instead of a borrowed password.

Payment rails built for agents have also shown how things go wrong. Security firm V12 recently found x402 payment flaws that could let a hostile API drain an agent's wallet, including no default spending cap. Stripe's own spending limits are still listed as coming soon, which makes the default human approval the main guard for now. Credential handling is a parallel problem, which is why 1Password now fills logins for agents without showing them passwords.

What is still missing

Several pieces are not available yet. Spending limits for agents and card creation are described as coming in the coming weeks or soon, without dates. Stripe's materials do not say how many businesses are in the preview, and the docs send readers to a separate availability page to check where Treasury itself is offered.

The approval model also shifts work rather than removing it. If every payment needs a person to confirm it, the time saved comes from the agent gathering data and preparing actions, not from paying bills on its own. Businesses that loosen the defaults for agent-tagged keys take on the risk of an agent being misled, for example by a fake invoice, and Stripe's documents do not describe extra fraud checks for that case.

AI Lookout did not test the feature. Everything above comes from Stripe's announcement, documentation and developer blog, read on 25 October 2026.

Who should try it

The natural users are small teams already on Stripe Treasury who want a daily cash summary or help preparing payouts without giving an agent the final say. A sensible start is read-only access: let the agent report balances and failed payments, then add write permissions one at a time.

Teams that build their own headless agents should create a separate agent-tagged key, keep Stripe's default approval rules, add amount thresholds and named reviewers, and wire up the approval webhooks so the agent knows when to stop waiting. Stripe's developer blog recommends exactly that layering: start with confirmation, then add approval rules once automation becomes part of daily operations.