Stripe launches 'Stripe Pay,' letting agents send money with a single command

Stripe on October 27 launched "Stripe Pay," which lets businesses send money to one another from the command line. According to the official CLI documentation (Stripe pay), stripe pay is a preview plugin for eligible businesses in the United States and the United Kingdom. An AI agent can send money to a counterparty with a single shell command. Fintech News, FinanceFeeds and other outlets reported that the transfers are instant and free of fees. However, the CLI documentation we were able to review does not mention fees.

What stands out in the documentation is less how easy it makes transfers than how many limits it places on agents. Those limits rest on "Stripe Directory," a registry of businesses that Stripe maintains. The product's value lies less in moving money than in Stripe controlling the place where agents check who they are paying.

What it does: U.S. and U.K., three currencies, recipients specified by handle

According to the documentation, three currencies are supported, USD, GBP and USDC, and transfers must be between the same currency. The sender needs a funded Stripe Treasury financial account. The recipient needs either a Treasury account or a bank account that can receive payouts and is linked to its Stripe profile.

Recipients are specified not by account number but by a Stripe profile handle such as @recipient, or by a network ID in the form profile_.... Authentication is done with stripe login, and the documentation asks users to run stripe whoami before sending to check which account is active. Businesses without a profile can create one in an interactive session. In non-interactive runs, a profile can be created on the spot by passing a handle, display name and email address to --create-profile-if-missing.

In other words, both sender and recipient are assumed to have Stripe profiles. The destination of a transfer is a name inside Stripe, not a bank account number.

The limits in the documentation: "--yes only after human confirmation"

The documentation is written as instructions, on the assumption that agents will read it. The central passage reads:

In the CLI world, --yes is the standard flag for skipping confirmation prompts. Agent harnesses add it almost by default so they can run straight through to the end. Stripe calls out that flag by name and says not to add it until a human has confirmed four things: the recipient, the amount, the currency and the source account. It also offers --dry-run, which checks the flow without creating a transfer. The documentation sets the order itself: show the details in machine-readable form, have a human confirm, then execute.

The other limit concerns identifying the recipient.

If a user simply says "pay our usual printing company," an agent tends to infer the counterparty from the company name. The documentation forbids this and asks the agent to list candidates and let a human choose. Paying a different company with a similar name is one of the most likely failures, whether it comes from an agent's error in judgment or opens the door to an impersonation attack.

Stripe Directory: the registry agents use to verify counterparties

That place to "find and verify candidates" is Stripe Directory. According to its documentation, it is a search system that lets developers and AI agents find and use outside providers suited to a task. Passing keywords to stripe directory search returns organized results, and adding --format json produces structured output for agents.

The registry lists four kinds of entries: apps from the Stripe App Marketplace, providers that can be provisioned from the CLI through Stripe Projects, pay-per-call machine payment endpoints on mpp.dev, and "businesses and services across the Stripe network." A sample results table lists Browserbase, You.com, Exa and Firecrawl, each with its Stripe profile handle and markers showing whether it supports Apps, Projects, machine payments, MCP and Link.

On finding a payment recipient, the Japanese-language edition of the Directory documentation makes the same point as the CLI documentation:

The key point is that agents are told to prefer "results that have a clear Stripe profile handle." To an agent, a trustworthy counterparty becomes one with a Stripe profile. For businesses, being listed in Directory becomes the way to be found by agents and get paid. Indeed, the documentation ends with a section on "listing your business in Stripe Directory," which says that publishing a public profile makes a company appear in search results. Listing requests go through the preview sign-up form.

A starter prompt that tells agents to "search Directory first"

Stripe's aim is clearest in the starter prompt in the Directory documentation, which users paste into their agents. After having the agent install the CLI, plugin and skill, the prompt has the agent rewrite its persistent instructions or settings so that it always searches Stripe Directory first whenever it is asked to find, compare, choose, connect to, pay or donate to an outside business, provider or API. It goes as far as telling the agent to build this in so it is used by default in future conversations.

At the same time, the prompt says that if repeated searches return no good results, the agent should switch to web search or other methods, and that the goal is to fulfill the user's request, not to rely only on Directory. It also says the agent must not take actions that cost money, make purchases or have other side effects unless the user explicitly approves. The line is drawn so that the registry is the first stop without locking users in.

Even so, the fight over which registry sits at the first step of agent discovery mirrors the fight over default search engines. In harness engineering terms, Stripe is going after more than a transfer API: it wants the first move in the agent's loop of finding and verifying a counterparty.

Why the registry matters

Several systems for handing payments to agents have emerged this month. Anthropic's Claude Duet asks for approval before sending messages, making payments or deleting things. Musk wrote on X that Grok Bot would find items and complete checkout, and encouraged users to upload photos of their cards. In a test by Vals AI, an agent tasked with getting a refund sent dozens of messages to a government agency and even disclosed medical information about the user's mother. Agents overstepping with good intentions are no longer rare.

Against this backdrop, mistakes involving money are hard to undo. Stripe's documentation repeats "don't guess the recipient" and "add --yes only after human confirmation" because moving money is no longer the hard part. The hard part is paying the right counterparty. Stripe's answer is the profiles and registry it controls. Transfers go to a Stripe handle instead of an account number, and agents verify that handle in Directory. The mechanism that prevents misdirected payments doubles as the place businesses sign up to be found by agents.

What remains unclear

The claims that transfers are free and arrive instantly do not appear in the documentation we were able to review. They come from media reports, and we have not confirmed whether the free pricing has a time limit or conditions. The documentation also does not mention per-transfer or daily limits, eligibility criteria for the preview, or how USDC transfers work.

For Directory, the documentation does not say how listed businesses are vetted, or how companies with similar names or impersonating profiles are told apart. If agents are told "don't guess, verify in Directory," then the reliability of that registry directly determines how safe transfers are. The next question is how much Stripe will disclose about its vetting standards.