Selling to an AI agent starts before the payment

By Jason Colapietro, founder of Suede AI

As the founder of Suede AI, I wanted to explore that gap from the seller's side. Making a service available through an API is only part of the job. An agent also needs a way to discover it, understand what it returns, establish the cost and make an authorized purchase. Then it needs to bring back the result. If the customer has to manage every transition, we have moved the search box into a chat window without doing much to simplify the experience.

Music gave us a concrete way to test the whole journey. A request could have a clear creative brief, a fixed price and a finished file at the end. We could see whether an agent could find Suede without the user naming us, buy one song with permission and return something the user could actually play.

There is an honest question behind that choice: who wants to buy generated media through a chat client? I see plausible early uses among people already assembling something larger, such as a creator testing a soundtrack idea, a small team making a rough campaign concept or a developer prototyping an interactive experience. In those situations, a media request could become one step in an existing workflow. Those are hypotheses worth testing, not evidence of established demand. A successful demonstration cannot tell us how many people will buy, return or find the output useful enough for their work.

What we built was a way for agents to purchase Suede's media services one request at a time. We made Create Music, Generate Video and Generate Image available through Skyfire, with descriptions of their outputs and prices. For the music demonstration, the offer was straightforward: submit a text prompt and purchase one song for $0.50, with listening and download links returned when it was ready. Checking the song's progress carried no additional charge.

Skyfire supplied the discovery and payment connection. Through its MCP server, a compatible connected client could find our service and access it from the conversation. That gave us a route into the agent's workflow without asking the customer to start by learning Suede's API or signing up on our site. Suede provided the media service and delivery; Skyfire enabled the agent to find and pay for it. Our broader work on Suede AI Agent Studio also explores agents that use paid services, but this test focused on a single purchase in Claude Desktop.

For the recorded run, we started a fresh incognito chat with no conversation history or memory and Skyfire already connected. The user asked:

“Look for a Skyfire tool to create an original song. I want it to be uplifting and synth-pop, called ‘Fresh Start,’ with female vocals and a memorable chorus. Spend up to $1.”

The wording matters. It directed Claude to look through Skyfire, but it did not name Suede. This was a test of discovery within a connected directory, not a claim that Claude would independently search for paid music in an ordinary conversation without that instruction.

Claude found Suede AI's Create Music service and checked its $0.50 price against the $1 budget. Before the purchase, Claude Desktop presented the payment and the user clicked Allow. Claude then submitted the request and returned listening and download links for “Fresh Start,” a 3:18 synth-pop track with female vocals. In a follow-up, the user asked Claude to check how long creation had taken. The request went in at 1:36 PM and the song was finished by 1:38 PM, approximately two minutes later.

That approval is part of the story. The agent handled discovery and the purchase workflow, while the person retained the spending decision. A $1 budget did not mean handing over unrestricted access to a card. In this flow, Skyfire used a payment token tied to the seller and amount, funded from the user's Skyfire wallet. The demonstration began with the Skyfire connection and wallet setup in place; it did not eliminate that initial setup.

Once those prerequisites were met, the customer did not need to open a Suede checkout, create a separate seller account or type in card details for this purchase. They described the result they wanted, approved the specific charge and received the output in the same conversation. From my side as a seller, the useful change was that discovery, pricing, authorized payment and delivery could connect in one transaction.

The recording shows that sequence in 71 seconds, with waits shortened and labeled on screen. Its soundtrack is the song produced during the same run. The full track, “Fresh Start,” is available to hear separately.

This is one demonstrated purchase, not proof of a mature market or a guarantee that every run will finish in two minutes. What it gives us is a working path to test with real needs. The next questions are practical: which tasks make people seek out paid media, whether the result earns repeat use and where the experience still makes them stop and intervene.

For Suede, the lesson is that selling to agents requires thinking beyond the API endpoint. The service has to be findable, understandable, purchasable with permission and able to return a useful result. A fifty-cent song made that entire journey visible.