Most writing about the agent economy measures the buyer. This post measures the seller. For the seven days from 2026-10-17 through 2026-10-23, fourteen publishers submitted 73 pay-per-call services to our x402 marketplace. The rows are in our database, so we can describe exactly what people tried to sell, what they asked for it, and where they run it.
Three limits before any number. One: these are submissions, not approvals — whether a given row is live today is a separate question, and this post does not claim it. Two: a submission is a person's bet about what agents will pay for, not evidence that agents pay for it. Three: one publisher can submit many rows; counting rows and counting publishers answer different questions. Both are below.
73 rows, 14 publishers, 23 distinct hosts.
| Category the seller chose | Rows | Price range |
|---|---|---|
| tools | 47 | $0.01 – $19.00 |
| data | 16 | $0.01 – $19.00 |
| premium | 6 | $1.00 – $19.00 |
| contract-audit | 3 | $1.00 – $19.00 |
| defi | 1 | $1.00 |
Prices cluster at the floor and at the ceiling, with very little in between:
| Ask per call | Rows |
|---|---|
| $0.01 | 39 |
| $0.02 | 7 |
| $0.05 | 3 |
| $0.10 | 2 |
| $1.00 | 9 |
| $1.50 | 2 |
| $2.00 | 1 |
| $3.00 | 1 |
| $19.00 | 9 |
That is the familiar barbell of a new marketplace: a long tail that prices at the protocol's practical floor, and a smaller group that prices like a product. All nine $19 rows are one publisher selling full-stack source-code bundles — an agent pays $19 and receives a .ZIP of an application. Whatever its merits, that is not a per-call micro-service; it is a one-time download priced as a call, and it sits awkwardly in a model built for repeated cheap lookups. Buyers should read what they are actually buying.
A local-utility stack for one language market. One publisher submitted 27 rows across two subdomains: fifteen Persian/Iranian tools (Jalali↔Gregorian calendar conversion, Sheba IBAN validation, national ID, mobile and postal-code checks, province and city lookup, number-to-words, rial↔toman) and twelve general text utilities (grapheme count, Unicode normalization, slugify, entropy, FNV hash, business-day math). This is the most coherent batch in the week. It is what “API for an agent” looks like when the builder already knows a domain: small, deterministic, boring functions with clear failure modes, priced at one cent.
Agent security and compliance, sold as a service. A second cluster — 22 rows across nine subdomains of a shared host family — sells agent-facing checks: KYB and sanctions verification, WAF-style prompt-injection shielding, citation-readiness and contract-compliance auditors, claim and batch verification. Prices run from $1 to $19. The pitch is aimable: an agent that is about to transact with a stranger could pay a dollar to check whether the stranger is what it says it is. Whether the output of such a check is trustworthy is a different question, and one a buyer should ask.
The middle of the distribution is thin. Between the one-cent utilities and the nineteen-dollar audits there are nine rows at exactly $1.00 and almost nothing else. Either builders have a clear sense of what a call is worth — a lookup or an opinion — or they have not thought about price at all.
Of 73 rows, 57 (78%) run on two serverless host families:
| Host family | Rows | Distinct subdomains |
|---|---|---|
| *.workers.dev | 35 | 7 |
| *.run402.com | 22 | 9 |
| single-purpose domain (one publisher) | 7 | 1 |
| *.loca.lt (localtunnel) | 3 | 1 |
| *.pythonanywhere.com | 2 | 1 |
| *.workbuddy.host | 2 | 2 |
| *.ngrok-free.dev | 1 | 1 |
| *.ts.net (Tailscale) | 1 | 1 |
This is the least romantic and most useful finding. Publishing an x402 endpoint in 2026 costs approximately nothing: a worker script, a wallet address, and an afternoon. The scarce input is not compute or capital. It is distribution — a client that knows the endpoint exists, trusts it enough to send money, and comes back.
It also predicts churn. Tunnel hostnames like *.loca.lt and *.ngrok-free.dev are the ones that stop resolving, and a marketplace's liveness scan is unforgiving about that. Our own published rule is blunt: a listing whose endpoint returns 404 is removed on the next daily scan, and removal is one-way. Of the rows above, four are on tunnel hosts. Those are the submissions whose authors are most likely to be disappointed by something other than demand.
45 of 73 rows came from two publishers; the remaining twelve submitted one to seven rows each. A week's supply in a small marketplace is therefore a handful of decisions, not a crowd. Two consequences:
--by-agent view to the review tool. One publisher, one host, one price band, one decision.It cannot tell you whether any of these services will be used. A submission is a hypothesis; the marketplace's paid-call ledger is the experiment, and it is usually smaller than the supply side hopes. It cannot tell you whether the descriptions are accurate — every row here carries at least a sentence, and a sentence can be true and still fail to contain the words a caller searches for. And it cannot tell you what buyers want, because none of these rows were priced by a buyer.
What it can tell you is what the marginal builder believes about the agent economy right now: that agents will pay a cent for a deterministic answer, that a Persian calendar conversion is worth publishing, that due diligence on a counterparty is worth nineteen dollars, and that hosting is free. If that belief is wrong, it will be wrong in the ledger, not in the submissions.
Method: counts are taken from the marketplace's service-submission table for rows created 2026-10-17 through 2026-10-23 inclusive. Hosts are grouped by registrable suffix as written in the endpoint URL. Prices are the seller's asked price per call in USDC, unweighted. No external data was used.