Providers
If we are measuring you
This page is for the people who operate the endpoints on this site. It says exactly what we do to them, at what rate, on what basis, and how to have a figure corrected, your rate reduced or your endpoints removed entirely.
Last updated 2026-08-16
01What we do to your endpoints
We call your publicly documented RPC endpoints on a schedule, time the response, check that it is well-formed and correct against the rest of the cohort, and store the result. Nothing else. The full method is published at /methodology, including every threshold and every window.
Specifically, and exhaustively:
- Read methods only. We never send a transaction, never sign anything, and hold no keys of any chain. Mutation methods are not benchmarked.
- Documented endpoints only. Every URL we call is one you publish for public use, or one you gave us a credential for. We do not probe for undocumented hosts, guess paths or scan ports.
- No attempt to exceed what you granted. We do not work around rate limits, rotate addresses to evade them, share credentials between endpoints, or use one account’s key against another’s host.
- No load testing. We measure what a normal caller experiences. Deliberately saturating an endpoint would tell us about your capacity and cost you money, and we do neither.
One further activity is not about your endpoints at all: we walk each chain’s public peer-to-peer discovery network to map where its nodes are hosted, using the same protocol those nodes run in order to find each other. We ask each node the one question discovery exists to answer and do not open a session to interrogate it further.
02At what rate, and how you can tell it is us
Conservatively, and identifiably. Default intervals put every endpoint well under one request per second, heavy methods run on much longer intervals than light ones, and each probe region schedules independently rather than in a synchronised burst.
Every request we make carries a User-Agent naming us and linking to our method:
RPCNode.IO-Probe/1.0 (+https://rpcnode.io/methodology)If you see that string in your logs, it is us, and the link goes to the page describing exactly what the request was for. We do not spoof a browser and we do not hide behind a generic client string.
An endpoint that refuses us is asked less, not more. When four requests in five over a rolling window come back rate-limited or plan-restricted, we drop that endpoint to one request every ten minutes until it recovers.
That backoff engages on refusal, never on failure. An endpoint returning errors is having an outage, and measuring outages is the point — only a rate limit or a paid-tier restriction counts, because those are the two answers that mean asking sooner will not help. It also never stops entirely: a quota that resets is picked up again within minutes.
03The basis on which we do it
We call endpoints you have published for public, unauthenticated use, in the manner you published them for, at a volume far below what an ordinary application would generate. A public RPC endpoint exists to answer requests from callers it has not met; ours are such callers, and they identify themselves, which most do not.
Where you require a credential, we hold one that we or a user of this platform obtained through your normal sign-up, and we use it within the plan it belongs to. We do not circumvent any technical access control, and we treat your published limits as limits rather than as targets.
The measurements themselves are our own observations of our own requests — facts about what happened on our side of the wire. Publishing them is reporting on a public service, and each figure is published with the window, the sample count and the method that produced it so that it can be checked or disputed rather than merely believed.
If your terms of service prohibit benchmarking or publishing performance data, tell us and we will stop measuring you and remove the published figures. We will not argue the point.
We would rather be removed from a provider’s estate than be in a dispute with them. The value of this platform is that its numbers are trustworthy, and a number obtained over an operator’s objection is not worth the argument it costs.
04What we promise about how you are presented
Ranking is never influenced by payment. There is no paid placement, no sponsored row, no arrangement under which a figure improves, and no advertising on any page that carries a measurement.
The composite score is one published formula with published weights, applied identically to every provider. It is computed from your own measurements against the cohort’s, and the components and their weights are on /methodology for anyone to reproduce.
Four rules protect against the ways a benchmark usually becomes unfair:
- Our own failures are never counted against you. A probe that could not reach the network, or a bug in our harness, is our outage. Those samples are removed before anything is judged.
- Rate limiting is reported as rate limiting. A provider enforcing its published limits is behaving correctly. It is a separate outcome from a failure and is never folded into an error rate.
- An unsupported method is a capability fact, not a failure. Choosing not to implement an optional method is a product decision, and is shown as one.
- Too little data is reported as too little data. Below the published sample threshold you get “insufficient data”, never a flattering score and never a failing one.
Where you publish an uptime commitment we quote it in your own words, and we state whether it covers the endpoint we actually measured — because a free public gateway is usually not covered by an SLA written for a paid product, and printing the promise beside the measurement without that would read as a company failing its own guarantee.
05What to ask us for, and what happens then
Four things, all of which we act on. Write to the providers address below from a domain we can tie to you, or from any address if you can point at something public that identifies you.
- A correction. If a figure is wrong, tell us which one and why. We keep the underlying observations, so we can check rather than take a view. If we are wrong we fix it and say what changed.
- A slower rate. Name an interval and we will use it. We do not ask for a reason.
- A different endpoint. If we are measuring the wrong host — a legacy gateway, a region you are retiring — tell us which one you would rather we used.
- Removal. We will stop measuring and unpublish. Removal is not conditional on giving a reason, and asking for it is not held against you anywhere on the site.
Right of reply: if you disagree with something published about you and we do not agree it is an error, we will publish your response alongside it in your own words.
One thing we will not do is quietly change a past figure. Where a correction changes what was published, the change is stated rather than applied silently — a measurement platform that edits its own history is not one anybody should rely on.
06Where to write
providers@rpcnode.io — corrections, rates, removal, right of reply
security@rpcnode.io — vulnerability reports
hello@rpcnode.io — anything else
A request that reaches the providers address is acted on before it is debated. If the ask is to slow down or stop, we do that first and discuss afterwards.