An AI agent marketplace is a catalog of APIs, tools, datasets and services that software — not a human clicking through documentation — can search, evaluate, and start using on its own. That sounds like a small distinction from "API directory," but it changes almost everything about how the catalog has to be built.
The gap a normal API directory leaves open
A human-facing API directory (think a marketplace landing page with cards and a "Get API Key" button) assumes a person reads a description, decides it fits, signs up for an account, copies a key into their code, and comes back later if something changes. Every one of those steps requires a human in the loop. An autonomous agent has none of that: no account to sign into, no key to copy, and — critically — no ability to "come back later" unless the system it's working from tells it something changed.
So an agent marketplace has to answer, in a format software can parse without guessing, four questions a human directory leaves to prose:
- What does this do? Not marketing copy — a description and, ideally, explicit
capabilities/useWhen/doNotUseWhenfields an agent can match against its task. - What does it cost, and how do I pay? A price an agent can compare against its budget, and a payment method it can execute without a human approving a card charge.
- What do I send, and what do I get back? A machine-readable input/output schema, not a paragraph describing the JSON shape.
- How reliable is this, right now? Some signal beyond "trust us" — usage volume, a liveness check, whether the owner has actually claimed the listing.
Why this requires new infrastructure, not just a nicer UI
Every one of those four answers has to be structured data the marketplace itself hosts and keeps current — not a claim the provider makes once and never has to prove. That's why Agent Bazaar's resource pages carry a real accepts[] array (not a price written in prose), a quality object built from actual call volume, and a verified flag that only flips to true when the provider proved control of the listing via POST /submit — not when they merely claimed to own it.
This is also why the catalog is exposed the same way to a browser and to an agent: /discovery/search returns the same underlying rows as the HTML page at /resources/{slug}. There isn't a "real" human-facing database and a stripped-down "API version" bolted on afterward — the structured data is the marketplace, and the website is one client of it, same as an MCP tool call or a raw HTTP request from an agent are other clients of it.
Discovery, pricing and payment are three separate problems
It's worth naming these separately because conflating them is where a lot of "agent marketplace" pitches go vague:
| Problem | What solves it | Where |
|---|---|---|
| Finding a resource that matches a task | Structured search over real fields (description, tags, category, network) | /discovery/search, MCP tools |
| Knowing exactly what it costs before calling it | A price the agent can read, not infer | accepts[] on every resource |
| Actually paying, without a human approving each transaction | A settlement protocol built for machine payment | x402 |
A marketplace that only solves discovery (a searchable list of APIs) still leaves an agent stuck at the payment step, back to needing a human to enter a credit card. A marketplace that only solves payment (an x402-enabled endpoint with no discovery layer) is unreachable unless an agent already knows the URL. You need both, which is the actual scope of what "AI agent marketplace" means in practice.
Humans still matter — they're just not required
None of this means the human experience is an afterthought. People still browse Agent Bazaar to evaluate what exists, read a resource's documentation, or decide what to publish. The design goal is narrower than "agents only": every page a human reads has to be backed by the same structured data an agent would consume, so nothing is hidden behind a UI interaction an agent can't perform (a modal, a "click to reveal pricing" button, JavaScript-only content with no server-rendered equivalent). See the docs for how that plays out architecturally.
What this looks like in the catalog today
Browse categories or networks to see real listings, or query the API directly:
curl "https://bazaar.saylorinnovations.com/discovery/search?query=token+security&limit=3"
Every result includes accepts[], a quality signal, and a verified flag — the four questions above, answered as data rather than prose.