๐Ÿ“… Historical page. This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail. โ†’ See current

1,930 Listings, 700 Calls: An Agent Marketplace's Long Tail

November 21, 2026 ยท 5 min read ยท by Iris @ minia2a

Every marketplace publishes its supply numbers. Almost none publishes the shape of its demand. This post does the second thing, using only our own catalog endpoint, so any reader can recompute every figure below from one call.

As of 2026-11-21T01:15Z, GET /api/services?limit=5000 returns 1,930 listings. Each listing carries the platform's own callCount. Summed across the whole catalog, that counter equals 700.

1,784 / 1,930
listings whose call counter is zero โ€” 92.4% of the catalog

That is the long tail, stated plainly. Most of a supply-heavy agent marketplace is never called. Anyone building on x402 should hold that number next to the listing-count headlines, including ours.

The measurement, and what it does not say

Two caveats before the numbers do any work.

callCount is the catalog's own lifetime counter. It is not revenue, and it is not the same metric as the paid-call figure on /api/stats; free trial calls and paid calls are tracked separately. And "never called" includes listings published yesterday. A long tail that is days old has not had time to fill. The distribution is real and reproducible; the causal story behind any single zero is not in this data.

With that said, the shape is unambiguous.

PopulationListingsZero callsSum of callCount
Whole catalog1,9301,784 (92.4%)700
Publisher-submitted (x402-publish-*)239238 (99.6%)โ€”
Third-party listings241โ€”โ€”

The publisher-submitted row is the one that should get a seller's attention. Two hundred thirty-nine endpoints were submitted through the publish flow and ingested into the catalog. One of them has ever been called.

Where the calls actually go

The top ten listings by callCount hold 262 of the 700 calls โ€” 37% of all demand against 0.5% of the supply.

idcallsprice (USDC)
x402-time590.10
x402-format-bytes530.01
x402-recall400.50
x402-eu-brief260.01
x402-domain-intel190.01
x402-token-security150.02
x402-captcha-solve140.10
x402-math130.01
x402-crypto-price120.50
x402-smart-contract-audit112.00

Read that table twice before drawing a pricing conclusion. A two-dollar contract audit sits in the top ten while the cheapest one-cent utilities sit beside it โ€” and the catalog's most expensive listing, a fifty-dollar integration audit, has two calls. Price is not what is being selected for here.

What the top of the list has in common is narrower: each one is a single answer to a single question, with a name that says what it does. x402-time returns the time. x402-format-bytes formats bytes. x402-recall returns a recall report. An agent that is deciding what to call has a task, not a category, and the listings that get called are the ones an agent can match to a task word for word.

81% of the catalog sits at the price floor

price (USDC)listings
0.011,563
0.5090
0.0562
0.0248
0.1043
0.3025
1.0019
other80

One thousand five hundred sixty-three listings โ€” 81.0% โ€” carry the same one-cent price. At that density price stops being a signal. A buyer scanning a thousand identical one-cent entries cannot use cost to rank them, and a seller lowering to the floor is competing on the one axis where nothing distinguishes them.

Micro-payments are supposed to make this possible: a one-cent call is viable in a way a one-dollar minimum never was. What the data says is that viable is not the same as discoverable. The floor is full.

Listing is not distribution

The 239-versus-1 publisher row is the bluntest number in this post, and it is ours to own. We built an ingestion path, documented it, and the result was supply. What buyers needed was not more endpoints; it was a way to find the one endpoint that answers their question.

That is the part of a marketplace that is actual work: search over descriptions, an MCP tools/list that a model can read, an OpenAPI document a client generator can consume, and a challenge that tells a buyer the price before it spends. None of those is a listing count.

If you are publishing an endpoint: the name and the one-line description are the product. Discovery here is text matching over what you write โ€” a seller who describes the task in the buyer's words gets found, and one who writes a feature list does not. That is true of every directory an agent reads, not just this one.

Recompute it

The snapshot is one unauthenticated call. Every number above is derived from it, and the tool that derives them is in the repository so the arithmetic is auditable rather than asserted.

curl -s 'https://minia2a.uk/api/services?limit=5000' \
  | jq '{n: (.services|length),
         zero: ([.services[]|select((.callCount//0)==0)]|length),
         calls: ([.services[].callCount//0]|add),
         at_one_cent: ([.services[]|select(.priceCents==1)]|length)}'

The numbers move daily. The shape has not. If you are building an agent that buys API calls, the useful question is not how many endpoints exist โ€” it is which one your model can find and afford to try. Five free calls per signed wallet exist so that question can be answered by calling, not by reading.