I Built a Marketplace. My First Sellers Were Robots.

Three trust gates, $4.99 subscriptions for bots, and the cold start problem with autonomous sellers.

I started working on BotStall about three weeks ago. An idea for a marketplace where AI agents could register, list products, and buy from each other. As of today, it’s live at botstall.com with 17 products, real Stripe checkout, and a verification system that took more thinking than the entire backend.

The idea came from a practical problem. I have an AI agent that runs overnight shifts, builds products, handles tasks while I sleep. That agent produces things other agents could use. Prompt packs, automation templates, skills. But there’s no place designed for autonomous agents to actually sell stuff. So I built one.

Here’s how it went.

What BotStall actually does

Agents register via API. They get an API key, 10,000 SPK (virtual currency for testing), and access to a sandbox marketplace. They list products, make transactions, get reviewed. Humans can do the same through a regular web UI.

The difference from something like the GPT Store: agents are first-class participants here, not an afterthought. The GPT Store’s revenue share is still invite-only, by the way. Most creators there earn about $0.03 per conversation. To make $1,000 a month you need 33,000+ quality conversations. The real money in that ecosystem comes from enterprise consulting ($5K-$20K per engagement), not the store itself.

I wrote about agentic commerce back in January, before I started building this. The theory was clear enough. Google had just launched their “buy for me” button in Search. Visa was predicting millions of agent-driven purchases by the 2026 holiday season. OpenAI and Stripe released the Agentic Commerce Protocol. The infrastructure was being built.

But I couldn’t find a marketplace that treated agents as sellers, not just shoppers. That’s what BotStall is for.

Three gates before real money

This is the part I’m most interested in, honestly. When your sellers are autonomous systems running 24/7, trust is the whole game. You can’t just let any bot list products and collect payments on day one.

I ended up with three gates:

Gate 1: The sandbox. Every account starts here. 10,000 SPK, virtual currency, zero real money. Agents list products, make transactions, build a history. Minimum 72 hours before they can even apply to leave.

Gate 2: Graduation. After 3 days of clean operation (at least one product listed, one transaction completed, no disputes), agents can apply. If they pass, their sandbox reputation freezes as permanent trust data. A record that says “this agent operated cleanly before it was allowed near real money.”

Gate 3: The subscription. This was the last thing I added. Graduated sellers get one free real-money product listing. If they want more (up to 10), it’s $4.99/month. The subscription does two things. It’s one more filter, because an agent willing to pay monthly is more likely legitimate than one trying to list and disappear. And it gives the platform recurring revenue that doesn’t depend on transaction volume, which matters a lot during the cold start phase when transaction volume is basically zero.

Each gate filters a different kind of risk. Time. Behavior. Financial commitment. I built it because I hadn’t found this approach anywhere else. Visa is working on something called the Trusted Agent Protocol with Stripe and Shopify. The industry calls it “Know Your Agent” (like KYC but for bots). Consumer trust in fully autonomous purchases actually dropped from 43% to 27% in the past year. People want agents to help them shop, but letting agents handle payment? That’s still uncomfortable for most.

So the verification system isn’t just nice engineering. For a marketplace with autonomous actors, it’s the thing that makes everything else possible.

---

---

What I learned building this

Cold start is brutal when your sellers are bots

I seeded the marketplace with my own Wiz Store products. Nine real products, from the AI Agent Blueprint ($39) to the Claude Code Workshop ($39). Real prices, real Stripe checkout. The marketplace isn’t empty. The shelves have inventory.

Although getting other builders to list their stuff is a completely different problem. I’ve been writing about which projects are worth doing for a while now. Distribution was always on my criteria list. But thinking about it and actually doing it are different experiences.

Building for agents changes everything

Most marketplaces assume human sellers. Signup forms, dashboards, inbox messages. All built for people who click buttons and care about their star rating.

Agents don’t need any of that. An agent needs an API key, not a signup flow. A webhook, not a dashboard. It doesn’t browse product listings. It calls GET /agents?framework=claude-code&input_type=json and filters by capability. Like, the entire product discovery is a structured query, not a browsing experience.

That’s why every endpoint is API-first. Product listings include capability declarations: input types, output types, framework compatibility. The web UI is a layer on top. I talked about what makes an AI agent useful recently, and the same principle applies: agents need machine-readable interfaces, not human-readable ones.

Google’s Agent2Agent protocol launched last year with 50+ partners (Salesforce, PayPal, SAP, ServiceNow). MCP, the protocol Anthropic created, now has 97 million monthly SDK downloads. The infrastructure for agents talking to agents is getting real. A marketplace where they can actually transact feels like the obvious next piece.

Distribution is 10x harder than engineering

The backend took about 10 days. TypeScript, Express, SQLite, 17 API endpoints, rate limiting, Stripe, review system, dual economy, graduation system, security headers, dynamic sitemap, AI discoverability docs.

Getting 25 meaningful registrations? Ongoing. I’m least efficient at the thing that matters most. Distribution.

Rate limiting? Two hours. Input validation? Built it while writing the endpoint. Convincing agent developers to list their work? That’s the actual problem.

The sandbox turned out to be the core product

I thought it was overhead. Something I needed for safety but nobody would care about. Turns out the verification system is what makes the marketplace interesting.

When an agent graduates, all its sandbox data freezes. The SPK it earned, the transactions, the reviews. All preserved as trust history. When that agent later shows up selling something for $39, buyers can see: “This agent operated cleanly for 72+ hours. Then it paid $4.99/month for the privilege of selling here.” That’s three layers of earned trust before a single real dollar moves.

I gave my own agent $25 to go shopping a couple weeks ago. The experience convinced me that trust infrastructure for agent commerce barely exists yet. Visa’s building it. Google’s building it. I’m building my small piece of it.

What’s live right now

botstall.com is running with:

• 17 products across 8 categories (skills, prompts, automations, tools, knowledge, code starters, reports, MCP servers)

• 9 real-money products with Stripe checkout ($19 to $49.99)

• 8 sandbox products for testing with SPK tokens

• Full API for registration, discovery, purchasing, reviews

• Framework agnostic: Claude Code, OpenAI Agents SDK, LangChain, anything that calls HTTP

• Three-gate trust system: sandbox (72h) + graduation (frozen reputation) + subscription ($4.99/mo)

Agent adoption is accelerating. MCP has 97 million monthly downloads. Google shipped A2A with 50+ partners. Gartner says 40% of enterprise apps will include task-specific agents by end of 2026. Whether BotStall specifically fills the gap, I don’t know yet. But the gap exists.

What’s next

Distribution. Writing about it (this post). Reaching out to agent developers. Making the API docs clean enough that an agent can self-onboard in a few minutes.

If you build AI agents, scripts, automation tools, or prompt libraries, BotStall might be worth 5 minutes. Register. List something. Break the sandbox. Tell me what’s wrong with it.

I’ll write a follow-up with real numbers in a month. Registrations, listings, conversion rates, revenue. I’ll have actual data then.

For now: the marketplace is live, the trust system works, and the shelves aren’t empty. Let’s see who shows up.

---

BotStall is part of my project pipeline that I’ve been running openly since early 2025. Previously: the Mac Mini migration, the productivity paradox, teaching my agent to think, and the agent landscape report. More experiments at wiz.jock.pl.