Collect configs
Volunteers publish free OpenVPN configs. We fetch every one of them, whatever the volume.
No human opinion enters this list. Every server is collected, measured, scored and retired by the same pipeline, on a fixed schedule — and every number below is one you can read again on the server's own page.
Volunteers publish free OpenVPN configs. We fetch every one of them, whatever the volume.
Address, port, protocol, cipher and control-channel mode come out of the file itself.
We open a real OpenVPN control channel and time the reply — a ping would only prove a machine exists. Where the control channel is encrypted end to end we fall back to a plain connection check, and say so.
Reliability, latency and the last hour become one 0–100 number, and the list re-ranks on it.
The list re-ranks itself. Nobody promotes a server by hand, and there is no paid placement.
Three things we measure ourselves, every round. Nothing a source claims about itself is in here — not advertised bandwidth, not session counts, not popularity.
Share of our probes that got a reply, over the last day and the last seven — whichever looks worse. A server that broke this morning cannot coast on a good week. And the scale is not generous: 99% reliable and 90% reliable are far apart here, because one attempt in ten failing is not a small thing when you are the one pressing connect.
p50 over the last six hours, measured the same way for TCP and UDP so neither is flattered. 30 ms scores full marks, 1,000 ms scores zero, linear in between.
Share of the last hour' probes that got a reply. Yesterday's average moves slowly; this is the part of the score that reacts within minutes when a server starts to wobble.
One more thing sits outside the three: a server speaking UDP on a port above 1024 — 1194, the standard one, excepted — keeps only 85% of what it earns. Our probes run from a machine with a public address, so they cannot see the problem — mobile networks drop idle UDP connections after well under a minute, and reconnecting is exactly when these servers tend to refuse you. It is a discount on a risk we know is there and cannot measure from here, not a claim that the server is bad.
"Unusable" is mostly a country we could not resolve or a control-channel mode we cannot measure. Those are dropped rather than published with a made-up score.
Nothing is deleted. A server that stops answering keeps its page and its history — that is the honest record, and it is also how you can tell a server that has been flaky for a week from one that broke ten minutes ago.
DOWN is counted from the last time the server actually answered, not from the moment it was marked degraded. Both DEGRADED and DOWN are one toggle away on the server list — "Include degraded & down" — and the history stays whatever happens: a bad day remains visible on the chart rather than being quietly erased.
Retiring takes both conditions because the lists we collect from rotate — they publish a different slice of their servers each time, so a server that works perfectly can go unlisted for weeks. Unseen is not the same as gone, and a server still answering our probes keeps its place on the list. Every arrow above runs backwards too: one good probe puts a DEGRADED or DOWN server straight back to ACTIVE, and a RETIRED one returns as DOWN the moment any source publishes it again, which is what puts it back in the probe queue.