The average agent-tool seller earns $26 a month
Machine traffic is now the majority of what hits this network, and a great deal of money is being raised on the premise that it can be billed per call. The public numbers do not support that premise. This is what four falsification passes found, with sources.
The headline numbers third-party measured
Per-seller figure derives from roughly $1.11M of 30-day volume across about 43,000 sellers, reported by an x402 operator using x402scan data, 2026-05-30. Artemis’ on-chain filter removed roughly 89% of headline x402 volume as non-organic. The conversion figure is one independent data marketplace reporting 1,183 agent probes, 5 settlements and $0.11 of revenue.
The interface itself is free, and that is deliberate eight of eight
Every major vendor that has opened a machine interface gives it away and meters the consumption underneath, which was already metered before agents existed.
| Vendor | What they say |
|---|---|
| AWS | “There is no additional charge for the AWS MCP server itself. You pay only for the AWS resources you create.” GA 2026-05-06 |
| Microsoft | Sentinel’s unified interface is “offered at no extra cost.” doc 2026-05-04 |
| Box | “available on all Box plans”; a tool call bills as one ordinary API call against the existing quota. read 2026-08 |
| Linear | “Linear’s stance has always been to make your Linear data accessible wherever you need it.” Framed as reach, not revenue. |
| Cloudflare | Has served billions of tool calls with no interface-specific charge. |
| An independent seller | Titled his own pricing post “Why We Charge $19/mo When the Market Average is $0.” 2026-04-23 |
Nine named incumbents shipped machine interfaces at MCP Demo Day in May 2025. Fifteen months later the published price is zero, and the reason is structural: after the 2026-07-28 spec made these servers stateless, exposing one costs about as much as a single worker process. Commodities do not hold price.
Per-fetch billing was tried by the best-placed operator and abandoned 2026-07-01
Cloudflare launched pay-per-crawl, ran it for twelve months, and bought a competitor in that market in January 2026. It then dropped the model entirely. The stated reason is about the unit, not the execution:
A single page might be crawled once and then cited in thousands of AI-generated answers, or crawled repeatedly and never cited at all. Cloudflare, on why per-crawl pricing was replaced with per-citation payouts, 2026-07-01
That is the central difficulty with per-interaction pricing for machine traffic, stated by the party with the most leverage and the best data. A fetch is not a unit of value. Over half of legitimate bot crawl traffic re-fetches pages that have not changed.
The buyers say no, in writing first-party
“It’s just an API; why do I need a wrapper around that when I can just call the API directly. This was my common response to these vendors as they tried to pitch usage of their MCP offerings (often at a premium!)” An engineer at Motion, 2026-03-14
The architecture is moving the same direction. Cloudflare’s own engineering post on code mode puts it bluntly: “Almost every MCP server is just a wrapper around an existing traditional API — why not expose those APIs?” Framework maintainers are following, and harness builders reported in August 2026 that they are migrating off tool-calling entirely.
Meanwhile the categories with real pull are saturated. One registry query returns 47 browser automation servers, 21 Postgres, 16 Slack — and it undercounts, because the two market leaders do not have the category word in their names. Those two draw 5,608,747 and 2,337,427 weekly npm downloads. Paid capability providers sit two orders of magnitude below them.
The payment layer is not the bottleneck evaluation is
The rails work. An agent can discover a priced endpoint, sign a payment payload, and settle, with no human present. That part is real, shipped, and audited. The problem is on the other side of the transaction.
The diagnosis from the marketplace that logged 1,183 probes and 5 settlements is that this is “evaluation friction, not payment friction: agents declining to pay because they cannot judge whether a resource is worth the cost.” A trust and scoring layer has been proposed and has no ratified specification. The protocol’s own payment proposal was opened in December 2025, never left draft, and both the issue and the pull request are closed. Payments do not appear anywhere in the roadmap published on 2026-08-22.
There is a second, quieter problem. At a median call price of $0.028, the metering costs more than it collects once you account for idempotent deduplication, per-agent caps, and sub-cent invoice aggregation.
Where the money demonstrably is for contrast
Per-interaction billing works, and it works in exactly one shape: a human with a budget buys a finished outcome, and the seller supplies the agent that produces it.
| Company | Model and result |
|---|---|
| Sierra | Outcome pricing. Reported $100M ARR in 21 months. |
| Intercom Fin | $0.99 per resolution, charged only when resolved. Reported growth from $1M to over $100M ARR on that model. |
| Zendesk | $1.50 per automated resolution. |
| Salesforce | Ran three pricing models in 18 months. Per-conversation pricing produced “tepid” adoption because buyers could not model their costs; the premium tier is now marketed on the words “Unmetered Agentforce usage.” |
Nobody in that table charges a foreign agent for access to anything. They sell completed work to a person with a purchase order. The distinction is the whole finding.
Method and its holes
Four independent research passes, each instructed to lead with the case against the thesis and to omit anything that could not be traced to a source read in full. Vendor claims are labelled separately from third-party measurement throughout, because most of the optimism in this space lives in the first category and most of the evidence lives in the second.
Known gaps, stated rather than papered over: no third-party agent-tool vendor publishes revenue, so the absence of a counter-example is weaker evidence than it looks. The x402 per-seller figure comes from a single operator’s snapshot and the seller count is inflated by wash-trading wallets counted on both sides. Reddit was unreachable to every pass, and a meaningful share of this conversation happens there.
Corrections are welcome and get published with the change logged: hello@crawlcensus.com.