Frequently asked questions

A relay (also called a gateway, proxy, or 中转站) is a third-party endpoint that forwards your requests to OpenAI, Anthropic, Google and others. People use them for regional access, pooled billing, cheaper pricing, or a single API key across providers.

The trade-off is trust: the relay sees your prompts and can swap the model behind the scenes. That is what this site measures.

We take no placement fees and no editorial sponsorships from relays. Rankings are the output of the formula on the methodology page and nothing paid changes them.

The outbound “Get this relay” buttons do go through kina.men, our click-tracker, which carries affiliate links to some relays. That is why the links are marked rel="sponsored". Click revenue helps cover probe hosting; it never affects ranking order or trust band.

Once a day at 06:00 UTC via a GitHub Action. The status strip at the top of every page shows the last scan time and the countdown to the next one. Traffic reports on the dashboard refresh every 5 minutes.

Yes. Every daily snapshot is committed to the public repo as data/snapshot-YYYY-MM-DD.json. A JSON API for the current snapshot is on the roadmap.

Because two of the four formula terms — model authenticity (40%) and price fairness (15%) — are still placeholders filled with a neutral 0.5. Under those placeholders the maximum achievable score is about 7.75, so nobody can hit the 8.0 “Trustworthy” floor until the authenticity probe ships. Today the spread you see is driven almost entirely by latency and uptime.

Evidence-based: a relay is labeled “real” when our probe actually observed real model identifiers at its /v1/models endpoint. Otherwise it is “unverified” — which is not an accusation. Many relays require authenticated requests to list models and so stay unverified until we have a probe key. The deeper authenticity probe that would distinguish real from substituted responses is still on the roadmap.

End-to-end time for a small completion request, currently from a single location (Singapore) via awesome-ai-api. Multi-region probes are in progress; until then a relay close to SG has a structural advantage, which is why latency is only 25% of the score.

Yes. Trust bands apply demotion flags on top of the score: an active substitution flag, an overpricing flag, or 3+ outages in 7 days will drop a relay to Risky regardless of its number. Flags never promote.

Where the relay primarily serves and bills. CN relays usually accept Alipay/WeChat Pay and are reachable from mainland China; GLOBAL relays bill in USD by card. It is a routing hint, not a quality signal.

A dashboard for up to 10 relays of your choice, weekly email reports, 90 days of history, and alerts when a watched relay changes trust band or has an outage. The public site stays free.

No. Subscribers get more visibility into the same data; nothing about a relay’s score or band is affected by who pays us.

From kina.men, which tracks clicks on the Visit buttons before they 302 to the relay. We record country, device class, browser family and referrer. No IPs are stored beyond the fingerprint window.

Use the Submit relay button in the nav — it opens the contact form with the listing topic preselected. We need a base URL, a models endpoint, and a probe key with a small balance. New relays sit in the Medium band for their first 14 days while history accrues.

Open an issue in the public repo with the snapshot date and the field you dispute. Because every snapshot is reproducible from its inputs, we can usually trace the number in a day. We do not adjust scores by request; we fix probes.

Yes, if you own the domain. Removed relays are listed as “delisted at operator request” for 30 days so users are not left guessing, then dropped.