verify[402]
snapshot · block — Inspect a 402

Field note · 8 min read · Nov 13, 2026

What x402's 402s actually look like.

300 public paid endpoints from the catalog's first page, one read-only probe each, no wallet: what a payment demand looks like when you actually read it.

Coinbase's x402 Bazaar catalog lists every paid endpoint that declared itself — 35,129 as of this week. We pulled the first 300 in the catalog's own order and read their 402s, using the same structural code an agent runs before signing anything. No wallet was involved. No payment was made. One probe per endpoint, 500 ms apart.

Scope, stated plainly

This is not an audit, not a red-team sweep, and not a reputation list. It is a measurement of what a payment demand looks like on the day we looked. Well-formed does not mean safe; your agent's policy is still what protects you.

The headline

ResultCountWhat it means
PASS222Well-formed challenge: network readable, asset named, price in raw units, payee named, resource bound, timeout within our default posture.
CAUTION60Well-formed, one flag a buyer should see — usually a long authorization window.
Refused (challenge_mismatch)9Header and body disagree — or advertise only parts of one another. The inspector refuses to guess; that is its behavior, not a verdict on the seller.
BLOCK1One challenge carried terms we cannot square with a sane payment: multiple payees and an amount we cannot represent.
Unreadable8Not an x402 challenge, not JSON, one timed out, four didn't resolve on the day.

282 of 300 were well-formed on read — 94%. That is the most useful number here, and the one we would not have guessed in July. It says the ceiling is high. The 6% that are not fine is where the work lives.

What the demand looks like

Chains. Across ~520 named accepts, Base (eip155:8453) carries 326, Solana 87, Arbitrum 41, Polygon 38, BSC 19, and everything else — Base Sepolia, X Layer, Algorand, Hyperliquid, a testnet or two — makes up the rest. Multi-chain challenges average about two accepts. If you support Base and Solana, you clear most of what is being sold.

Tokens. Every priced challenge we could read resolves to canonical USDC — 283 of them, all the same token at the same decimals across Base, Solana, and Arbitrum. Stablecoin USDC is the price of admission on this shelf; other dollar tokens did not clear the list.

Prices. Median $0.005 a call. p90 is $0.02. The largest request is $1,000. Menus are wide, adults are cheap.

Methods. 209 are GET, 91 are POST. Half the priced endpoints would answer a plain GET probe.

Volume. Across the sample, the catalog records 344,634 calls in the last 30 days — median 84 calls per endpoint, with a long tail dominated by a handful of endpoints doing 10–80k a month. One of them is mirroring our own price exactly (challenge 2000 USDC minor units vs our 50000). An uneven sample, but it is what the public said it used.

The finding worth acting on

56 challenges ask for an authorization window of one hour (maxTimeoutSeconds: 3600), and 57 endpoints are flagged for it under a 15-minute posture. That does not make those sellers evil; it means the authorization you sign sits usable for an hour, and the reasons that window exists are commoditized convenience, not malice. Long-time listeners will notice the same thing we noticed: this is exactly the spot where a caller's own policy does the stopping — policy.maxTimeoutSeconds: 900, and the 57 over-long entries come back CAUTION to your agent before it signs.

Names (well-formed only, by real volume)

Not endorsements: names are given only because the demand is public data and the endpoint passed the same structural check we apply before any money moves. All resolve to Base USDC and all clean-parse.

Endpoint30d callsPrice
blockrun.ai/api/v1/chat/completions80,319$0.002
stableenrich.dev/api/exa/search34,871$0.010
api.agents-trust.com/recommend16,441$0.010
agent.pocket.network/v1/meadow15,494$0.005
api.exa.ai/search12,810$0.007 (CAUTION)
stableenrich.dev/api/exa/contents8,527$0.002
glim.sh/api/v1/twitter/search6,237$0.005 (Solana)
api.kadec0.xyz/v1/token-price6,186$0.020
www.api-xpay.com/x402/ping5,787$0.005

Unnamed endpoints matter more than my handful of names: the 222 PASS and the 282 well-formed total are above the fold of anyone's catalog. Read yours if you run one; the check is free and it does not sign anything.

What we refused — and what that means

Nine endpoints got a challenge_mismatch. Some of those are dual-dialect twins (same demand in x402 v1 and x402 v2 to hedge client generations). Some advertise only part of what the header advertises — a subset. In both cases we refuse to compute a verdict instead of silently picking one; refusing is not an accusation, it is the correct behavior for a tool meant to pick the demand out of the noise of what a server says.

The one BLOCK had multiple_payees and an amount that does not survive representation. We do not publish which; the file carries the verdict and the reasons, not the name.

The part where this report did its job on itself

This scan ran twice before it read well. On the way, it caught two defects in our own inspector: (1) it compared the wrapper resource and refused real endpoints that echoed their challenge in both header and body; (2) it compared challenges dialect-naively and refused v1-body/v2-header twins of the same demand. Both fixed, tested, and live, in commits 41586f6 and b788b32. A measurement pass that does not improve its own instrument tells you nothing; we got a better instrument out of this exercise and so did the seller whose endpoint was refused the first time.

The fine print

The same read now runs daily and is published as a public board — every row timestamped, filterable, with an embeddable badge for the seller: /board, machine-readable at /board.json. The raw JSON for this one-off snapshot is in the repository. The method, the caveats, and the tool were published when the report was; nothing in it was selected on outcome. If your endpoint is listed and you disagree with our read of it — that is expected; check it yourself before anyone pays, and tell us what we got wrong. That is the best kind of finding, and the only kind we asked for.

The ask

Read one yourself.

Paste any public 402 into /inspect. If it comes back CAUTION or BLOCK on a field you did not expect, that is worth knowing — before your buyer does.

Previous: What verify402-mcp 1.1.0 ships. · All notes