Stripe is moving its role in payments further upstream.
In demos shown at Stripe Tour, agents were not limited to completing a payment. They were shown building products, calling outside services, creating orders and then finishing transactions after a user approved them. Taken together, the demos point to a broader goal: Stripe is not just adding a pay button to chatbots. It is trying to connect agents to real commercial systems.
From a prompt to a live website
In one demo, Annice Liu, a solutions architect for Stripe Greater China, asked Kimi to launch a family travel service. The service would generate a personalized weekly travel newsletter for parents based on their preferences and the latest local information.
The agent then began assembling the required services on its own. It registered a domain, deployed the site through Cloudflare, connected to models through OpenRouter and used Supabase to store user preferences. Within minutes, an idea that had existed only in conversation became a working website.
The demo also illustrated how the overseas software ecosystem is structured. Rather than placing an agent inside one closed platform with every function already bundled together, the model is to let it combine a set of independent services.
Cloudflare handled the domain and deployment. OpenRouter provided access across models. Supabase supplied the database. Stripe Projects tied activation, configuration and payment together. Each company focused on one layer, and the agent assembled them into a finished product based on the task at hand.
Stripe co-founder and CEO Patrick Collison had previously said Stripe Projects was inspired by how Claude Code works. Once an agent writes a product, it quickly needs logs, hosting, databases and model access. Stripe wants to make those services discoverable and directly provisionable by agents.
According to Collison, Stripe Projects went from idea to public release in about two months with two or three engineers, integrating roughly 50 services along the way. He estimated that in the past, the same work would have required a larger team and at least half a year.
Open composition versus the super app model
In China, a more common path starts inside a super app. Accounts, payments, merchants, content and local services already sit inside one ecosystem, making it easier for an agent to complete the loop from user request to transaction. The trade-off is different. The flow is shorter and users do not need to manage multiple accounts, but what the agent can call and how data moves are more tightly bounded by a single platform.
The open model used overseas is not automatically easier. Every new service an agent connects to introduces another set of identity, authorization, billing and risk questions. Developers used to handle those one by one. If agents are going to build products themselves, they need an infrastructure layer that turns services into resources that can be discovered, activated, metered and paid for.
That framing also offers another way to look at Stripe’s acquisition of OpenRouter. In the demo, OpenRouter looked like a model interface that spared the agent from integrating with each model provider separately. For Stripe, though, it also represents one of the most frequent forms of consumption in AI products. Every model call involves trade-offs across price, speed, reliability and output quality, and each call creates a measurable token cost.
If model routing, token metering, billing and payments are linked together, Stripe is no longer managing only how a product gets paid for after it is sold. It is also involved in how an AI product chooses and purchases model capacity while it is running.
From that angle, Stripe is trying to provide a shared layer for money, metering and permissions across an open ecosystem, so agents can move across multiple companies while still getting something closer to the continuity users expect inside a super app.
Meta’s Muse shows the consumer side
Another demo featured Meta’s personal agent Muse. A user asked it to find a round-trip flight from Boston to Dublin. Muse found an option priced at $470, confirmed the passenger details and then asked whether it should complete the booking.
At the payment step, it did not ask the user to hand card details to the AI. Stripe’s agent wallet, Link, surfaced an authorization screen. The user could see the amount, the card being used and a choice between Deny and Allow. Only after approval did the agent receive the payment credential needed to complete that specific transaction and book the ticket.
Placed side by side, the two demos read like a compact picture of Stripe’s view of the AI economy. On one side, agents can build products, call services and generate revenue. On the other, they can search for goods, create orders and initiate payment on a user’s behalf.
Stripe’s aim is larger than enabling a chatbot to pay or get paid. It is trying to connect agents, as software actors, to the commercial systems where transactions actually happen.
Turning products and checkout into agent-readable interfaces
The first problem in AI-assisted shopping is not how to charge a card. It is how an agent can accurately understand what a merchant is selling.
Humans can open a webpage, inspect images, filter sizes, understand shipping and then click through checkout. If an agent can only imitate a human by hunting for buttons on a page, it can easily choose the wrong variant, read an outdated price or break when the page changes.
In 2025, Stripe and OpenAI introduced the Agentic Commerce Protocol, or ACP. It turns product catalogs, inventory, pricing and checkout into a machine-callable flow. The first public use case at the time was Instant Checkout inside ChatGPT.
The significance was not simply that users could buy things inside ChatGPT. Stripe had started turning checkout from a page built for humans into a commercial interface that agents could call directly.
Stripe also launched Agentic Commerce Suite. The idea is to let merchants integrate once and then serve multiple AI entry points, without rebuilding product and checkout systems separately for ChatGPT, Gemini, Copilot or the next agent platform.
Merchants still remain the merchant of record. They continue to control inventory, taxes, shipping, refunds, chargebacks and customer relationships. AI platforms bring the request and the order. Stripe translates the natural-language request back into the merchant’s existing transaction system.
That is also where Stripe differs from pure AI platforms. OpenAI, Google and Meta are competing for the user entry point. Stripe wants to be the commercial infrastructure behind whichever entry points win. It does not need to predict the final winner among agents. It needs more agents to participate in transactions.
Muse follows a similar path. It first uses browser actions so merchants without dedicated integrations can still be reached. It then connects directly to important partners through APIs for more speed and stability. Mark Zuckerberg’s view, as cited in the article, is that companies in the future will have their own agents in the same way they now have websites, email addresses and social accounts.
Browser control answers whether the system can work now. Direct interfaces answer whether it can support stable commerce later. That is why Stripe is trying to reshape products, checkout and payment capabilities into agent interfaces.
Writing limits into the payment credential itself
Once an agent can create an order, the next question is simple: on what basis is it allowed to use a user’s money?
Traditional online payments often treat possession of the correct card number as a key trust signal. For agents, that is not enough. Holding a payment credential does not mean the agent has authorization for the transaction in front of it.
Stripe’s answer is Shared Payment Token, or SPT. Instead of handing a real card number to the agent, Stripe generates a restricted payment credential. It can be limited by merchant, amount, currency and expiration time, and it can be revoked. The agent can use it to pay, but it cannot turn it into a general-purpose card that works everywhere.
Stripe showed a clear example at Sessions 2026. A token was issued with a $25 limit. A $25 charge succeeded. When the merchant changed the amount to $50, Stripe declined the transaction.
That may look like a simple spending cap, but the article describes it as one of the most important design choices in agentic payments.
If the budget exists only in a prompt, the system is merely asking the agent to behave. If the limit is embedded in the payment credential, the payment system can still block the transaction even if the model makes a bad judgment or is influenced by malicious instructions.
Large language models interpret ambiguous intent. Payment systems enforce hard rules.
This year, Stripe also expanded SPT to Visa and Mastercard agentic tokens, as well as buy-now-pay-later methods such as Affirm and Klarna. The company is not trying to build a new payment network from scratch. It is trying to translate agent authorization across different networks into payment credentials merchants can already accept.
Link is becoming an agent wallet
SPT addresses how merchants can safely receive money. Consumers still need a place to manage which agents can ask them to pay, what they approved and how much was ultimately spent.
That is where Link Agent Wallet fits.
An agent can send Link a spend request that specifies the merchant, amount and purchase purpose. The user reviews it on the web or on a phone. After approval, Link returns either a single-use virtual card or an SPT to the agent. The user’s underlying card details are not exposed to the AI.
The Muse flight booking demo is one example. Stripe also showed a smaller one at Sessions: Claude wrote a joke-filled letter for the presenter’s daughter, then used PostalForm to mail the PDF. Link displayed a $3.40 request on the phone. The user saw the merchant and purpose, approved it with Face ID, and the agent completed payment and mailing.
One case involved airfare. The other involved postage. The amounts were different, but the logic was the same. Users are not handing their wallet to AI on a standing basis. They are approving a specific task with context, amount and recipient.
Each Link Agent Wallet payment still requires user confirmation on a transaction-by-transaction basis, according to the article. Budget rules and approval-free spending may come later. By September, Muse, Grok Bot and Instinct were already using Link Agent Wallet. Agentic purchases completed through Link had also grown 38x in one month.
The article adds an important qualifier: growth is fast, but absolute transaction volume is likely still at an early stage.
From a demo to actual use of Muse
The author says that when the Muse demo first appeared at Stripe Tour, they had not used the product. More recently, they started using Muse for short-video work and for planning travel in the United States. The process, in their description, felt smooth. Muse did more than list suggestions. It remembered earlier instructions and continued searching and organizing later steps. On that basis, the article argues that personal agents may indeed become a new consumer entry point.
Meta’s commercialization plan for Muse has also become clearer. Zuckerberg said the base version of Muse would remain as free as possible, while Muse Wonder would be monetized through subscriptions. Over a longer horizon, Meta hopes to take a small share of the value Muse creates when it helps users save money, make money or complete transactions.
Zuckerberg also gave a specific example. He said he barely used Instacart before, but registered an account so Muse could buy ingredients for cooking based on a conversation. He now uses it once or twice a week. His point was that agents may do more than improve conversion. They may create transactions that would not have happened otherwise. In time, account relationships may no longer belong only to humans. Agents may gradually have accounts of their own.
The article is careful not to overstate the ease of that business model. If merchants pay commissions, Muse has to answer whether it recommended a hotel or product because it was the best fit for the user or because the commission was higher. If users are charged directly, Muse has to prove that the time and money it saves are worth the fee. Then there are mistakes, price changes, cancellations, refunds and after-sales support. Any weak point can erode trust in the agent.
The article also separates Meta’s economics from Stripe’s. Muse is closer to a transaction entry point, seeking subscription or conversion revenue from helping users get things done. Stripe provides payment credentials, authorization, issuing, risk controls and transaction processing. It is trying to make money from infrastructure.
Stripe’s consumer relationship is moving forward from checkout
For Stripe, that creates a different kind of opportunity. Acquiring payments already means handling consumer transactions, and Link has long stored consumer payment methods directly. In the past, though, Stripe’s relationship with consumers mostly appeared inside a merchant’s checkout page. The user paid, and Stripe helped the merchant collect the money.
Agent Wallet moves that relationship one step earlier. Stripe starts managing the connection between consumers and agents directly: which agent can request payment, what it can buy, how much it can spend, whether another confirmation is needed and what protections exist if something goes wrong.
The shift is that Stripe is no longer only processing a consumer payment. It is also participating in how consumers delegate payment authority to agents.
If personal agents do become the place where people search, compare and buy services, Stripe does not need to build the most popular agent itself. It only needs to become the wallet that different agents want to integrate and that consumers are willing to trust. That would move Stripe from being the processor behind merchant checkout to being infrastructure called when consumers authorize agents to spend.
The article also notes how difficult that path is. Handling merchant-initiated collection and continuously managing consumer authorization for agents are not the same product relationship. The latter brings brand, experience, privacy and constantly shifting user trust into the equation. Still, if Link Agent Wallet can secure the payment entry point for personal agents, the upside would be larger than simply launching another payment product.
Letting machines buy APIs, data and services directly
When a consumer asks an agent to buy a flight, the transaction still looks like familiar e-commerce. But agents will also become buyers in their own right.
An agent may need weather data, image generation, model inference, identity verification or another agent’s capability for a single task. Those purchases are often worth only a few cents or a few dollars, and they can happen at very high frequency.
Traditional API purchasing flows are not built for machines. Registering an account, applying for an API key, choosing a monthly plan and binding a card usually require human involvement. An agent that wants to query data once should not have to buy a year-long subscription first.
Stripe and Tempo introduced Machine Payments Protocol, or MPP, to address that. It uses HTTP 402, Payment Required, so a service can tell an agent in the same request that the resource costs money, how much it costs and how it can be paid for. The agent then pays using existing authorization and receives the result.
The article cites two demos. In one, an agent created and deployed an API service, and another agent used MPP and Link to pay $2 for a code review. In another, an agent bought weather data for a very small amount. Payment could be made with cards carried through SPT or with stablecoins.
This line of business is distinct from AI shopping on behalf of people.
In the first case, the agent represents a human buyer, so identity, authorization, after-sales support and liability matter most. In the second, software is buying resources to complete a task, so the priorities are programmability, micropayments, observability and automated reconciliation.
Stripe is pursuing both because it believes agents will become not only a new shopping interface but also a new class of enterprise customer and software buyer.
Collison’s list of next questions is concrete: how agents should register for and use Stripe, whether MCP is sufficient, how Stripe CLI should adapt, and how agents should pay one another and in what currency. His view is that within roughly three years, measured by transaction count, most transactions may involve agents on both sides. That does not mean agents will account for most transaction value. A more likely outcome is a very large number of small agent-to-agent payments.
That would turn payments from a discrete purchase action into a base protocol for agent collaboration. To complete a task, one agent may repeatedly call models, data, software and other agents. Every call needs identity, authorization, billing and settlement behind it.
Paying is only part of the problem
Real transactions do not end when money moves.
Airfares can rise at the last step because of baggage fees. Cards can trigger 3D Secure. Items can go out of stock. Agents can buy the wrong size. Users can ask for returns. Merchants also need to decide whether they are dealing with a trusted agent or an automated attack pretending to be one.
Recent updates to Link Agent Wallet begin to address those issues directly.
If the final price rises because of taxes, shipping or add-on services, the agent can request an increase to the authorized amount instead of placing a new order and freezing another payment. If extra verification is required, Link can give the agent a link that returns the user to a 3D Secure or identity-check flow. If the original card is declined, the agent can ask the user to switch cards and submit a new spend request.
At the same time, the system has started allowing agents, with user permission, to read spending history to infer frequent merchants, budgets and product preferences. That could make personal assistants more useful, but it also raises new questions: how much financial data should a shopping agent see, can it use past spending habits to shape future choices, and do users really understand what they have opened up?
Purchase protection is another notable piece. Eligible Link agent transactions now cover accidental damage, lost items, price drops and fee-free returns, according to the article. Muse is the first agent to integrate that protection.
That suggests the real competition in agentic payments is no longer just about who enables the first AI-driven payment. It is about who takes responsibility when the AI gets the purchase wrong.
In payments, the hardest thing to replicate has never been the button alone. It is the authorization, risk controls, refunds, disputes and liability allocation behind the button.
The four interfaces Stripe is trying to own
Looking back across Stripe’s moves over the past year, the article groups them into four positions:
- how agents discover and understand products;
- how merchants receive orders coming from agents;
- how users safely delegate part of their spending authority to agents;
- how machines directly buy APIs, data and compute resources.
ACP and Agentic Commerce Suite address the first two. SPT and Link Agent Wallet address the third. MPP addresses the fourth. Radar, Checkout, Issuing, Financial Connections and Stripe’s existing refunds and dispute systems reconnect those new entry points to the traditional payments world.
Taken together, the picture fills in from both ends. Zuckerberg’s view is that agents will gradually have accounts of their own and create new consumption. Collison’s view is that many transactions may eventually have agents on both sides. The article explicitly says these are directional judgments, not established market facts. Even so, they help explain why Stripe is building across merchant interfaces, personal authorization and machine payments at the same time.
Stripe is not simply trying to let AI use Stripe. It wants any AI platform entering commerce to rely on the layer Stripe already understands best: turning a vague purchase intent into a transaction that merchants can accept, users can revoke and systems can hold accountable.
Agents can touch payment credentials, but they are not fully autonomous spenders
The article closes with a clear caution. This is not yet a mature stage where AI can simply spend on its own.
Agents today can do more than generate shopping lists, create orders or bring users to checkout. They can, after user approval, obtain restricted payment credentials and continue the transaction flow.
That is still far from saying AI can independently control a user’s money.
Most public examples still rely on per-transaction approval. Approval-free budgets remain part of future planning rather than a default feature. Cross-platform refunds, duplicate orders, identity misuse and liability assignment are also far from fully solved.
For that reason, the article suggests 2026 may be remembered as the first year of protocols and infrastructure for agentic payments. What is changing now is that agents have entered transaction decision-making and started touching payment credentials, while payment companies are rewriting forms of trust that used to be implicit in human commerce into rules machines can execute.
A mature version of agentic payments probably will not look like handing AI a card and letting it improvise.
It is more likely to look like this: users define goals and boundaries first, agents search, compare and execute, and wallets and payment networks perform a final check on amount, merchant, timing and purpose. If something goes wrong, there is a clear chain for refunds, disputes and responsibility.
On a large screen in Shanghai, that may look like a smooth flight booking. The article’s point is that what Stripe is really selling is the reason that transaction can be safely approved in the first place.
The original article also states that its content reflects personal observation and thinking only, and does not constitute investment or business advice.


