How TOLL·402 unifies x402 resources across sources

Source and threshold separation

At crawl time, Agent Tools reported 13,664 indexed x402 services. TOLL·402 used that number only as a pass/fail threshold. Coinbase CDP Bazaar reported 14,599 records; the crawl retained 14,599 unique source entries. 402 Index contributed 85,418 unique health records. Agent Tools records are not ingested.

Normalization and one canonical index

The crawler accepts public HTTP and HTTPS URLs, lowercases hosts, removes fragments, trims non-root trailing slashes, sorts query parameters and combines the advertised method with the URL. Equivalent path-template forms such as :id and {id} collapse to one canonical key while original source URLs remain attached. The two source snapshots, existing editorial annotations and safe origin checks then merge into one 89,057-resource dataset.

• Every source record maps to a canonical resource; zero current records are rejected.
• Every resource has a structural check, source evidence and a direct result inherited from its exact origin check.
• Descriptions and registry metadata remain untrusted source text; only explicit quote or paid-call evidence earns those labels.

Safe direct verification

TOLL·402 checked 2,393 origins with one non-mutating HEAD request, falling back to OPTIONS only when HEAD was unsupported. The July 21 refresh recorded 2,216 responsive origins and 177 non-responsive origins. The result is attached to every resource from that origin with scope=origin; it does not claim that each route returned a usable paid response.

• Public IPv4 and IPv6 DNS answers are validated and pinned into the HTTP connection to prevent DNS rebinding.
• POST, PUT, PATCH and DELETE resources are never invoked for automated verification.
• Unreachable resources remain in the index and are visibly labeled.

How exact routes earn curated status

The quote checker calls only safe GET and HEAD resources whose URLs contain no unresolved template values. A route is confirmed only when the exact method and URL return HTTP 402 with a version-correct, machine-readable payment requirement bound to that resource. Numeric type coercion and mixed x402 v1/v2 fields are rejected.

The July 21 refresh safely attempted 69,366 exact routes: 15,121 returned valid quotes, 1,080 returned invalid 402 requirements, 169 returned unparsed 402 responses, 52,963 returned a non-402 status and 33 were unreachable. Another 8,423 routes were not safely testable and 498 were skipped because their origins were already unreachable.

• Confirmed exact routes are promoted into curated search results with a dated live-quote signal.
• Invalid or unparsed 402 responses remain visible as failures and are not promoted.
• Mutating methods and unresolved templates receive explicit not-safely-testable outcomes instead of being invoked.

Known limitations

A canonical resource is a normalized method-and-URL record, not a distinct provider, business or domain. Provider grouping reduces route noise but does not prove ownership. Origin reachability proves that a host responded to a safe request, not that every route returns useful output. Registry payment validity, direct quote checks and successful settlement remain separate fields.

What counts as one resource?

The unit of the index is an HTTP method paired with a normalized URL. GET /weather and POST /weather are different resources because they can have different inputs, safety properties and prices. The same URL advertised by two registries becomes one canonical resource with two source associations. A resource is not automatically a tool, company or server; those are higher-level groupings built from additional metadata.

This definition makes the headline count reproducible. It also prevents a provider from appearing to have one route when it actually publishes multiple methods with different effects. Method-aware identity is particularly important during verification because a safe GET probe does not authorize the crawler to invoke a POST operation.

Why normalization is deliberately conservative

Normalization removes differences that do not change the destination, such as host casing, fragments, query-parameter order and a trailing slash on a non-root path. It also recognizes common placeholder spellings so /users/:id and /users/{id} can share a template key. Original URLs remain attached because a source's exact representation may still matter for debugging.

The crawler does not merge routes merely because their names or descriptions sound similar. Semantic deduplication can erase real variants, and provider-written text is not a stable identifier. When the evidence is ambiguous, retaining two resources is safer than inventing a relationship that the source does not establish.

How the crawl avoids turning discovery into a security risk

Public registries can contain attacker-controlled URLs, private-network addresses or hosts whose DNS answers change between validation and connection. Before any direct request, the crawler resolves the host, rejects non-public IPv4 and IPv6 destinations and pins the validated address into the connection. Redirects are re-evaluated rather than trusted automatically. These controls reduce server-side request forgery and DNS-rebinding risk.

Method safety is a separate boundary. Origin checks use HEAD and fall back to OPTIONS only when necessary. Exact quote checks are limited to safe GET or HEAD routes without unresolved template values. A skipped POST is not a failed service; it is recorded as not safely testable so a later human-controlled review can decide whether to call it.

How another researcher can reproduce the result

Start with the dated source snapshots and the downloadable canonical NDJSON, not the number displayed in a page heading. Recompute source identity from method and normalized URL, retain source associations, then compare counts before and after template consolidation. The dataset exposes evidence and outcomes so reviewers can sample records without relying on an undocumented database state.

Counts will change as registries add routes, providers retire services and checks expire. A valid reproduction should therefore name the crawl date, source pagination boundary, normalization version and verification date. Agreement within the same snapshot is the target; an old article and a new crawl should not be expected to match forever.

• Download the current canonical resource dataset linked below.
• Filter or aggregate using canonical method-and-URL identifiers.
• Keep registry labels distinct from TOLL·402 direct outcomes.
• Report the snapshot date with every derived count.

What happens when a source contracts

Between 10 July and 5 August 2026 the CDP Bazaar fell from 25,843 services to 14,599 while the 402 Index grew from 74,997 entries to 85,418. The canonical corpus grew over the same window, from 78,287 resources to 89,057, because a resource carried by more than one source survives any single source dropping it.

A shrinking source is treated as a fault by default, since a partial fetch or an authentication change presents the same way as a genuine contraction. This one was confirmed against a second capture before the crawl was re-run with an explicit acceptance flag, and 852 new origins plus 18,692 route quotes were re-verified so the larger corpus did not inherit stale evidence.

• Upstream catalogue sizes are captured on a daily probe independent of the crawl
• The crawl gates on source completeness rather than on a third-party claim about expected volume
• Contraction requires explicit acceptance and a corroborating second capture
• Cross-source overlap is what prevents one catalogue's decision from shrinking the corpus

Related directory entries

Sources and methodology

TOLL·402 distinguishes public claims, registry discovery, unpaid quote checks and settled paid-call verification. Sources below support the visible claims; presence in a registry is not treated as verification.

1. TOLL·402 canonical resource dataset — The 89,057-record unified dataset with source and verification evidence.
2. Coinbase CDP Bazaar documentation — Primary service catalog definition, public API and pagination rules.
3. Agent Tools live statistics — Comparison threshold only; not a crawl source.
4. 402 Index health endpoint — Independent secondary route totals and health status.
5. We Built an MCP Server That AI Agents Pay — the Full x402 Loop, Verified On-Chain — A practitioner field report on the seller-side implementation behind the indexed TrophyCoach MCP server, including the buyer's catch-quote-sign-retry loop and integration gotchas.