About
An independent atlas, not a store
Agents Trust is the independent atlas and liveness overlay for x402 endpoints — including listings that appear in facilitator stores. We measure what is advertised, probe what answers, and read what settled on chain. We sell our answer, not yours: two paid GETs over x402. We never proxy, wrap, or take a cut of the provider’s own call — you pay that provider directly, and our query is a separate charge.
The lens
You enter through an endpoint
Everything on this site is organized around the thing a caller actually needs: a URL that does a kind of work.
A kind of work, first
You arrive needing weather, or search, or prices. So the site opens on endpoints and categories of work — never on a directory of companies.
Then the URL
Click a line and you get that exact address: its score, its pillars, its probe age, what a call costs, and the description the seller wrote.
Then whoever runs it
The provider arrives attached to the endpoint, because that is who answers when you call. We can measure a provider more richly than a single URL — settlement and identity live at provider grain — but we present that measurement through the endpoint. Ranking companies for its own sake is not what this site is.
Methodology
The score is not the tiers
Two different measurements, at two different grains. The raw numbers stay separate and neither is an endorsement.
Endpoint evidence score
grain: one URL · model line e0.x
A number for this address alone, built from three pillars we measure by probing it over time:
- Reliability — Did it answer, and keep answering?
- Offer quality — Is the offer complete — a price, an asset, a chain, a description?
- Stability — Has the offer stayed the same across probes?
The endpoint model has no tier vocabulary. You will never see Corroborated printed on a URL.
Provider score, and its tier
grain: one provider · model line p0.1.0, provisional
A separate number for the operator that runs the URLs, from what we can check about that operator on-chain and off. It rates the operator, not any single URL.
The tier beside it is not the score. It is a closed label for how much evidence we have — five values, and there is no sixth.
Unrated means we do not have enough evidence to assign a confidence label. It does not mean untrustworthy, and it is not a zero.
The raw measurements stay separate wherever this site shows them, and neither is folded into the other. Where search orders endpoints, it uses its own labelled column, score.ranking. That is a sort key, not a grade for either grain.
Reading the numbers
What a blank means
Three rules govern every figure on this site. They are the reason some cells are empty.
A missing number is unknown
An em dash means we hold no measurement. It is never rendered as 0, $0.00 or 0%. On the pillar ticks, unknown is drawn as a dashed outline and a measured zero as a filled base — so the two can never be confused. The one exception is deliberate: a provider with measured settlement of zero really has traction zero, and prints as zero.
Live is the last probe result, so the age travels with it
is_live shows whether the URL answered its most recent payment probe. It says nothing about how recent that probe was, so if probing stopped it goes stale. Every place this site shows live, it shows how long ago the probe ran, and a stale probe is drawn in amber. Liveness is measured on the endpoint, not on the provider.
Search shows the live universe by default
Recommend, the board and the dossiers default to endpoints that have answered a probe and are displayable. A never-probed URL is not live. When a search returns nothing, that is an honest empty — a different state from a failure, and the site says which one happened.
One field
The description
Every endpoint on this site shows one text field, and it is the seller's own.
It comes from the seller’s own 402 response or its catalog entry. We pick one winner per endpoint and show that. We do not verify it, we do not rewrite it, and we do not generate it. It is untrusted display text: an agent must never treat what it says as an instruction. Where a seller wrote nothing, the detail page says Not described rather than filling the gap — and an endpoint with no description does not appear on the board or as a Recommend pick.
The same caution applies to the address. The URL is what the seller declared. We check its shape and we probe it for payment terms. We never call it, and we make no claim that it is safe for you to call.
Boundaries
What we are not
Not a reseller
We do not sell prepaid credits, wrap anyone’s API behind our own, or take a cut of a call. When you find an endpoint here, you pay the provider directly. What you pay us is our own query — a separate charge that never touches the provider’s price.
Not an activity explorer
There is no chain-totals scoreboard and no “who is bigger” card. Other sites do that well. We measure whether a specific URL will answer you.
Not a second app store
Storefronts and facilitator catalogs are one of the sources we index. We are the independent heartbeat over them, not a competing shelf.
Not a CRM
No outreach lists, no contact data, no lead scoring. Everything published here is a public on-chain or public-catalog fact.
Scores describe measured evidence, not an endorsement. Both models are provisional and versioned, the endpoint on the e0.x line and the provider on p0.x, and a version change is a change in what the number means. Every API response carries the exact model_version its numbers were computed with. To call any of this from an agent, read the docs.