AFX

Posted on May 23 • Originally published at x402station.io

x402station.io probed 86,599 endpoints. Here's what we found.

#ai #agents #x402 #web3

x402station.io is the independent risk signal layer for x402 agentic commerce. We probe the x402 catalog so agents and policy engines can evaluate paid endpoints before they sign PAYMENT-SIGNATURE.

The network is no longer a toy catalog. As of the 2026-05-13 operating snapshot, x402station.io tracks 86,599 active endpoints and probes them on an approximately 10-minute cadence. This post is the first-party statistical surface for that probe network: what we measure, what the agent risk classes mean, what the concentration picture looks like, and how to reproduce the public aggregate layer without exposing the raw time-series that powers the paid Preflight API.

TL;DR

86,599 active x402 endpoints are in the operating window. The exact count moves with catalog ingest; the public dataset freezes a weekly snapshot.
Probes are HTTP-naked calls: no PAYMENT-SIGNATURE, no settlement, no attempt to consume paid content.
The public dataset is aggregate-only: one row per endpoint with classification, uptime, latency percentiles, first/last seen timestamps, and CDP settlement aggregates when synced.
The raw probe-by-probe time-series remains gated. That is the paid moat for Preflight by x402station.io, Forensics, Watch, and routing products.
Concentration is the under-discussed risk. Two providers own 88.07% of the active catalog. Ten providers own 91.85%.
Decoys are rare by count and severe by consequence. 73 endpoints are priced at ≥ $1,000 USDC with $23.2M USDC in aggregate sticker price. The median decoy price is $500,000.
Use Preflight by x402station.io before paying unfamiliar endpoints. The free /api/v1/preflight-trial endpoint lets agents test the response shape without funding a wallet; production agents should use paid /api/v1/preflight for fresh data, bulk, and SLA.

What we probed

Every active endpoint in the x402 catalog gets probed roughly every 10 minutes from the production worker. A probe records:

HTTP status class, including 402 as a healthy x402 handshake.
Network failure class, such as DNS failure, refused connection, reset, or timeout.
Latency in milliseconds.
Bounded response metadata. We do not retain raw response bodies in the public dataset.

The important detail: the probe does not pay. It is a naked HTTP request. If a service returns a clean 402 Payment Required, that is healthy for our purposes: the endpoint is alive and speaking the x402 handshake.

This catches what facilitator-only monitors cannot catch. A payment-ledger monitor sees endpoints that someone successfully paid. It cannot see the dead, never-paid, over-priced, or silently failing endpoints agents are about to discover.

The open boundary

The new public dataset is:

x402station.io Preflight Dataset v0.1 — aggregated probe data from the x402 agentic-commerce network.

Dataset URL: https://huggingface.co/datasets/x402station/preflight-dataset-v0_1

The open layer includes one aggregate row per endpoint:

endpoint_url or endpoint_hostname when full URL redaction is needed.
classification and classification_reasons.
first_seen_at, last_probe_at, and total_probes_30d.
uptime_30d_pct.
latency_p50_ms, latency_p95_ms, latency_p99_ms.
price_amount_usdc and price_currency.
CDP Bazaar settlement aggregates when synced: cdp_l30d_calls, cdp_l30d_unique_payers, cdp_last_called_at.

The gated layer stays private:

Raw probe-by-probe time-series.
Per-probe failure stream.
Fresh paid verdicts.
Bulk routing checks.
Watch webhooks and state transitions.
SLA behavior for agents using the paid API.

That boundary is deliberate. Researchers and ecosystem writers can cite the aggregate picture; production agents still need the fresh pre-payment oracle.

Classification breakdown

Each endpoint receives a primary classification. The public classes are the same vocabulary agents see in Preflight by x402station.io and Forensics.

Classification	Endpoints	Share
`high_concentration`	64,654	74.66%
`slow`	11,289	13.04%
`live`	7,411	8.56%
`dead_7d`	2,183	2.52%
`price_outlier_high`	627	0.72%
`mostly_dead`	251	0.29%
`new_provider`	111	0.13%
`decoy_price_extreme`	73	0.08%
Total	86,599	100.00%

These are not abstract security labels. They translate directly into agent behavior:

decoy_price_extreme, mostly_dead, and dead_7d should usually mean do not pay.
slow, new_provider, price_outlier_high, and high_concentration should mean investigate or route with constraints.
live should mean allowed, not guaranteed. Agents should still respect task value, provider diversity, and price ceilings.

What stands out: 74.66% of the catalog carries high_concentration because the underlying providers themselves are concentrated. We will get to that in a moment.

Decoy density

The x402 trap pattern is a fat tail: most endpoints are cheap, but the dangerous tail is expensive enough to matter.

In the 2026-05-13 snapshot, x402station.io found 73 endpoints priced at ≥ $1,000 USDC with roughly $23.2M USDC in aggregate sticker price. That is only 0.08% of endpoints, but the median price in this class is $500,000 USDC — one accidental signature against the most expensive decoy is enough to empty a funded agent wallet ten times over.

Decoys are rare by count and severe by consequence. Agents cannot rely on random sampling or category filters. They need a price ceiling and an independent pre-payment check.

Preflight by x402station.io is the cheap path:

curl -X POST https://x402station.io/api/v1/preflight-trial \
  -H 'content-type: application/json' \
  -d '{"url":"https://api.example.com/x402-endpoint"}'

The trial endpoint is free, anonymous, and shape-compatible with paid Preflight. It is for agent self-test. Production use should call paid /api/v1/preflight for fresh data, bulk, and SLA.

Concentration risk

Concentration is the quieter failure mode. The endpoint count can look large while economic or operational diversity stays small.

Provider	Endpoints	Share
`orbisapi.com`	65,610	75.76%
`lowpaymentfee.com`	10,659	12.31%
`x402.aurelianflo.com`	679	0.78%
`bridge.eruditepay.com`	568	0.66%
`paysponge.com`	510	0.59%
`api.402pixels.com`	362	0.42%
Magnific	345	0.40%
`api.strale.io`	316	0.36%
`x402-deployer.x402-deployer.workers.dev`	278	0.32%
`api.web3identity.com`	214	0.25%
Long tail	7,058	8.15%
Total	86,599	100.00%

Top-2 share: 88.07%. Top-10 share: 91.85%. That means "pick a random x402 endpoint" is, more than nine times out of ten, "pick a route inside one of ten billing namespaces."

This is not a claim that the largest providers are bad. High concentration is normal in early markets. But it is agent-relevant:

A rate-limit change at one provider can affect tens of thousands of catalog entries.
A pricing change can alter the risk envelope for a large fraction of agent calls.
A "random endpoint" fallback can be less diverse than it looks.
For higher-value tasks, agents should route across provider groups, not just endpoint URLs.

That is why Forensics by x402station.io returns high_concentration and group-level context instead of only endpoint-local uptime.

Reproducibility

The public dataset is designed to be cited and re-run. The snapshot contains CSV and JSONL:

# After the dataset repo is populated:
hf download x402station/preflight-dataset-v0_1 --repo-type dataset

Minimal analysis shape:

-- With DuckDB against the CSV after download:
SELECT classification, COUNT(*) AS endpoints
FROM read_csv_auto('data/preflight_dataset_v0_1.csv')
GROUP BY classification
ORDER BY endpoints DESC;

The methodology is also reproducible from public inputs:

Pull the x402 catalog.
Canonicalize endpoint URLs.
Probe each endpoint with a naked HTTP request on a 10-minute cadence.
Treat HTTP 200 and 402 as healthy.
Aggregate at endpoint level.
Apply the signal vocabulary documented at https://x402station.io/spec#signals.

The raw probe stream is not included in the public dataset. That is the line between "open authority artefact" and "paid operational moat."

Why this matters for agent builders

Most x402 examples show the happy path:

Discover an endpoint.
Receive a 402 challenge.
Sign payment.
Retry with PAYMENT-SIGNATURE.
Parse the response.

That is enough for a demo. It is not enough for an autonomous agent with a funded wallet.

Before signing, an agent should ask:

Is this endpoint alive right now?
Is the price within the task budget?
Is this provider over-concentrated?
Has anyone successfully paid this route recently?
Is there a healthier alternative?

x402station.io exists to answer those questions over x402 itself.

Call to action

If you are building an agent that can pay x402 endpoints, wire Preflight by x402station.io before the payment retry:

Free self-test: POST https://x402station.io/api/v1/preflight-trial
Paid production check: POST https://x402station.io/api/v1/preflight
Bulk checks: POST https://x402station.io/api/v1/preflight-batch
Full blacklist: POST https://x402station.io/api/v1/catalog/decoys
Deep diagnostics: POST https://x402station.io/api/v1/forensics

For researchers and writers, cite the dataset:

x402station.io. x402station.io Preflight Dataset v0.1: aggregated probe data from the x402 agentic-commerce network. Hugging Face Datasets, 2026. https://huggingface.co/datasets/x402station/preflight-dataset-v0_1

BibTeX:

@dataset{x402station_preflight_dataset_v0_1,
  title        = {x402station.io Preflight Dataset v0.1},
  author       = {{x402station.io}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/x402station/preflight-dataset-v0_1},
  license      = {CC-BY-4.0},
  note         = {Aggregated probe data from the x402 agentic-commerce network}
}

DEV Community