What We Actually Built: Selling Machinery an Agent Can Operate

Sam Bakker · Founder, gocushy · July 20, 2026

A payment link takes money once. That's the whole job it was built for, and it does that job well. But a payment link isn't a business. A business needs to sell a little more on each order, to bill again next month, to pay the people who send it customers, to charge the right tax and issue a real invoice, to hand money back cleanly when something goes wrong, and to chase a card that just got declined. That collection of unglamorous, load-bearing parts is what we call the selling machinery — and until recently it only existed inside dashboards a human had to sit down and click through.

This is the "here's what we actually shipped" post. Over the last months we built the machinery as a set of tools an AI agent can operate directly, through the Model Context Protocol, on your own Stripe account. Twenty-eight tools in total. You connect an agent through the npm package @gocushy/mcp (Claude Desktop, Claude Code, Cursor) or a remote MCP endpoint (ChatGPT and other hosted agents), and there's a plain REST API and an ordinary human dashboard alongside. Every payment settles with you — Stripe Connect direct charges — so gocushy never holds your funds and never becomes the merchant of record. Here's what's inside, grouped the way it actually gets used.

Capture more per sale

Three mechanics do most of the work on average order value, and your agent can wire all three from a sentence. Order bumps add a checkbox offer under the card field. One-click post-purchase upsells charge the card the buyer already gave you, no re-entry. Coupon codes let you run a launch price or a partner discount without rebuilding the offer. "Take the $49 course, add the $17 workbook as a bump, and set a 20%-off code for my list" is a handful of tool calls, not an afternoon in a settings screen.

Make it recurring — and keep it paid

Subscriptions bill on your Stripe, with renewals generating their own orders and receipts. Recurring revenue has a failure mode, though: cards expire and get declined. So the machinery includes email-platform connectors to multiple ESPs and built-in follow-up and dunning that chases failed payments instead of quietly writing them off. And when a sale completes, delivery webhooks fire so whatever you use to grant access hears about it immediately.

Grow through other people

You get your own affiliate program, not a bolt-on. It tracks referrals, holds commissions back before they come due, and claws them back automatically on refunds — the part most home-grown affiliate setups quietly get wrong. Your agent can create an affiliate, pull their link for a specific offer, and manage the payout workflow without you opening a screen.

Stay compliant

Selling across borders drags tax and paperwork in whether you planned for it or not. Stripe Tax runs against your own registrations. Every sale can produce a sequential, compliant invoice. Refunds work from both the dashboard and the API and stay synced with Stripe in both directions, so the two systems never disagree about what was returned. (This is product information, not tax or legal advice — where you must register is a question for your accountant.)

Run it without holding your breath

Operating a store through an agent means you want a clear record and a way to rehearse. Every action — by an AI, a human, or the system itself — is written to an account activity log. There's two-factor auth on the account. A test-drive mode lets you walk your own checkout end to end with the payment simulated, so you see exactly what a buyer sees before a real card is involved. And there's a public status page, because a platform that touches money should be honest about its uptime in public.

The honest objection: "you're going to let an AI touch my store?"

It's the right question, and the answer isn't "trust the model." It's that we built the platform so the agent cannot do the damaging thing, whatever it's told. Concretely:

• Prices live server-side. The record set when a product is created is what a buyer is charged. An agent can create an offer; it literally cannot misprice a product.
• Checkout copy is restricted to whitelisted blocks. The agent fills fixed slots — headline, bullets, badge, button text. It can't rearrange the layout or ship a broken page.
• Countdown timers are honest by construction. They hide at zero and cannot be made to fake-reset. An agent told to restart the timer for every visitor will fail, correctly.
• Buyer-facing pages use signed URLs. Delivery pages and invoices can't be guessed or forged — by an agent or anyone else.
• No silent dead ends. The platform warns the agent when a product has no delivery method set, so nothing goes on sale with no way to fulfil it.

How this got built

Worth being straight about, since it's the same bet we're asking you to make: the large majority of gocushy's own code was written by AI agents, and then reviewed and adversarially attacked by other AI agents before any change was accepted. Underneath that sits 330+ automated tests and multiple security review passes. And we dogfood it — gocushy's own founding checkout runs on gocushy. None of that makes the software perfect. It does mean the guardrails above aren't a marketing diagram; they're the constraints we run our own money through.

That's the whole pitch, really. Not a prettier cart — AI made pretty pages effectively free. The machinery a business actually runs on, rebuilt so your agent can operate it on your own Stripe, safely.

Frequently asked questions

Sam Bakker is the founder of gocushy. He has a long track record building launch and funnel software for creators, and builds from New Zealand.