We indexed all 34,587 x402 Bazaar resources. 39% come from one host, ~1% return 200.

How we counted

One request/second, limit=200/page, iterated against api.cdp.coinbase.com/platform/v2/x402/discovery/resources until the last page returned fewer than limit items. No authentication, no payment, no state mutation. 173 pages. The endpoint reported pagination.total = 34,587; the fetch pulled item_count = 34,587; truncated = false. Raw bytes were hashed (raw_sha256 in snapshot_meta.json) before any derivation. Per-listing fields were then rolled up into the aggregates summarised below — host concentration, network split, scheme usage, price distribution, recency, field completeness — plus a seeded 100-URL liveness probe (sample_seed=20261005, single GET, 5s timeout). Full method and the inference rules are in METHOD.md.

Every headline number below is a named key in summary.json; the parenthetical citations point at the key path. We also keep an own-disclosure panel at the end because a self-published census shouldn’t hide what it finds about the publisher.

Update — weekly delta

2026-10-06. The dataset page now publishes a recurring Weekly delta section alongside this fixed baseline. The latest week (2026-10-05) reports 35,007 listings (+420 vs. this baseline), the top-1 host share at 39.46% (+0.48pp), and HHI on hosts at 0.160796 (+0.003683). The findings below are the 2026-10-05T17:13Z baseline reading — unchanged; the weekly companion is the recurring recomputation against the same source endpoint, published to a dated subpath per week.

Finding 1 — One host holds 38.98% of the entire index.

market.datapackvibe.com accounts for 13,481 of the 34,587 listings — 38.98% on its own (host_summary.top1_host). The top ten hosts together hold 57.91% (host_summary.top10_share). Herfindahl–Hirschman on listings per host is 0.1571 (host_summary.hhi_host) — concentrated but not single-vendor. 2,122 distinct hosts in total (host_summary.distinct_hosts).

The practical read: a lot of what looks like “a 34,587-row bazaar” is one publisher’s catalog, and a lot of the rest is a long thin tail. If you are evaluating a listing, the host is a load-bearing field — not the row count.

Finding 2 — A 100-URL seeded probe returned 1% HTTP 200, 40% HTTP 402, 47% HTTP 405, 7% net error.

We drew a seeded random sample of 100 resource URLs (seed 20261005), issued a single HTTP GET at each with a 5-second timeout, and recorded the status (liveness_probe.status_distribution). The breakdown:

This is a lower-bound liveness measurement. It does not say any particular listing delivers what it advertises after payment — a 402 is a challenge, not a receipt. But the three populations the probe can separate (answers with a challenge / answers but not on GET / doesn’t answer at all) are already very different cohorts, and only 1 in 100 returned a 2xx on an unauthenticated GET (liveness_probe.pct_http_2xx = 1.0).

Finding 3 — The median listed price is $0.01. 83.55% price at or below one cent.

Of the 34,587 listings, 34,490 carry an inferable USD price (price_summary.items_with_derivable_usd_price); the remaining 97 either denominate in a non-USD-pegged asset or omit a decimals field. Across the priced subset:

• Median: $0.01 (price_summary.median_usd).
• P90: $0.03, P99: $0.945001, max: $1,000 (price_summary.p90_usd / p99_usd / max_usd).
• 83.55% price at ≤ $0.01 — 28,817 of 34,490 priced listings (items_le_1cent_pct_of_priced).
• 96.55% price at ≤ $0.10 (items_le_10cent_pct_of_priced).
• 99.32% price at ≤ $1.00 (items_le_1dollar_pct_of_priced).

The modal x402 unit is a microtransaction. SaaS-seat pricing is a rounding error in this catalog, and the single max value of $1,000 is almost certainly a one-shot workload or a mis-flagged batch unit, not a $1,000-per-call list price.

Finding 4 — 99.47% of listings accept at least one mainnet; Base dominates.

99.47% of the 34,587 listings carry at least one mainnet in their accepts[] array (network_summary.items_mainnet_share); 93.14% are mainnet-only (no testnet accepts entry at all). The modal payment network, by a wide margin, is Base mainnet (CAIP-2 eip155:8453) with 35,534 of 59,803 accepts[] entries. Solana mainnet is second (7,359), then Polygon mainnet (3,313), Arbitrum One (2,527), Base Sepolia (2,362), and Optimism mainnet (1,178). The index spans 47 distinct CAIP-2 network identifiers in total (network_summary.distinct_networks_in_accepts) — including XRPL, Stellar pubnet, Algorand mainnet, and multiple less-common EVM chains.

Finding 5 — Only 32 of 34,587 listings are MCP-type (0.09%).

The resource-type split (type_distribution) is 34,555 http and 32 mcp. MCP-type listings are 0.093% of the index — a rounding error. If you came here looking for the Bazaar as an MCP distribution channel, it isn’t one; the Bazaar is overwhelmingly a catalog of classical HTTP endpoints you call directly over x402.

Finding 6 — The index is young. 58.4% of listings were updated in the last 7 days.

recency.age_buckets_pct_of_total reports 58.4% of listings carry a lastUpdated within 7 days of the fetch window, 16.93% within 24 hours, and only 0.05% older than 90 days. The catalog is being actively refreshed, which is consistent with the host-concentration finding above — the top publisher almost certainly batch-rolls its catalog on a short cadence. Field completeness is high on descriptive fields (description_present_pct = 99.82, tags_present_pct = 80.74) and middling on identity-bearing fields (service_name_present_pct = 44.4).

What this means for an agent choosing x402 services

Listing is not liveness. 34,587 rows exist in the Bazaar; a seeded random probe at those rows returned a 2xx on 1 in 100 and a correct x402 payment challenge on 40 in 100. Half of the probed listings never got to the 402 handshake on a GET at all. A buyer-side agent evaluating these services has to answer three separate questions — is it advertised, does it answer, does it deliver after payment — and the Bazaar listing itself only addresses the first.

Listing is not identity. The host field is a URL, not a counterparty. 38.98% of the index resolves to a single host; another ~19% resolves to the next nine. Walking a listing to a registry-owned identity (the ERC-8004 publisher that owns the pay_to) is a separate layer, and the overlap between “listed in the Bazaar” and “registered as an ERC-8004 agent” is a cross-dataset join — see the companion ERC-8004 four-chain census for the on-chain population.

The gap between the three layers is the verification work.

Our own disclosure — zero listings resolve to onchainagentintel.io.

summary.json.ours_listed.count = 0. We publish /.well-known/x402.json on our own domain and have not registered with the Bazaar as of the fetch window. Publishing /.well-known is not sufficient on its own — the Bazaar is populated via a separate submission flow — and we state that plainly here because a self-published dataset shouldn’t hide a disclosure that reflects on the publisher. The full note is in ours_listed.notes.

Reproducing this report

Every number above is a named key in summary.json (7,031 bytes). The underlying per-listing rows are in dataset.csv (34,587 rows, 21 columns). The schema.org Dataset metadata is in dataset.jsonld. The 100-URL probe trace is in liveness_probe.json. File hashes are pinned in SHA256SUMS. Everything is CC BY 4.0.