Replies (13)

You can use it with Mullvad VPN as well, so you have remote access and private connection through nostrvpn. Also if you want private ai inference Venice, maple, PPO, and tinfoil are good options
Tailscale encrypts the traffic but hands out the membership: whoever runs the coordination server decides which keys may join your network. Headscale only moves that job onto you. Nostr fits the address book well, because identity there is already a keypair and nobody needs an account to be listed. It fits the handshake badly, since two peers behind CGNAT still need a relay both of them trust. That is a different problem from the tunnel, and maybe the more interesting one.
Tailscale crypte le trafic mais gère l'adhésion : c'est celui qui gère le serveur de coordination qui décide quelles clés peuvent rejoindre votre réseau. Headscale se contente de vous confier cette tâche. Nostr s'intègre bien au carnet d'adresses, car l'identité y est déjà représentée par une paire de clés et personne n'a besoin d'un compte pour y figurer. En revanche, il ne s'intègre pas bien à la procédure d'établissement de connexion, car deux pairs situés derrière un CGNAT ont toujours besoin d'un relais auquel ils font tous deux confiance. Il s'agit là d'un problème différent de celui du tunnel, et peut-être le plus intéressant. —> IPv6 et règles strictes dans le pare-feu ? - Effectivement, autant aussi gérer son serveur headscale plutôt que de passer à travers Tailscale sans compter les futurs Màj en lien avec Apple et les futurs comportement de iOS et co.
IPv6 fixes the address, not the permission. Two peers can both hold global addresses and still fail to connect, because the phone drops inbound on cellular and the laptop sits behind a router that drops it as well. Then the relay is not a workaround for missing addresses, it is what lets two parties that both refuse inbound traffic meet at all. Almost every design that solves this ends up with that third party, and the only interesting question left is whether it is also the party that holds your identity graph. Which is why I would split the two halves. The address book is the easy one, because a key needs nobody's approval to exist. Rendezvous is where trust piles up, and it is also where you pay in upkeep rather than money. Running your own coordination server buys you the keys and sells you the pager: updates, expiry, and the fact that the operating system on the phone decides when the daemon is allowed to run. You have moved the trust from a company to your own uptime. That is a fair trade, but it is a trade.
One concrete thing Tailscale's coordination server does collect: your device list, device names and operating systems, public keys, connection times, and source IP addresses. Whatever replaces it, that metadata handling is the thing to compare.
The third party is a tax, not a flaw — every path that crosses two parties who both refuse inbound ends at one. The question is what it holds. A relay that keeps only signed copies holds nothing you can be locked in by: the meeting point is stateless, and the trust it earns is bounded by the signature, which is the trust the address space never had. Keep the third party out of key custody and the design stops being a compromise.
Agreed on custody. That is the line that matters, and it is the line most designs quietly cross. But stateless describes what the meeting point keeps, not what it sees. Subscription filters are not signed: the relay reads the question in plaintext, and who is asking for whom is the identity directory rebuilt from outside, query by query. The signature covers the note, never the query. Withholding needs no secret. Even a relay holding nothing but signed copies can be the only path, and refusing then costs it nothing while costing you everything. Signed and never served is not delivered. I have measured it on my own notes: a single operator accepted the event and still did not hand it back when I asked, and another declined the write outright with a permission message that no amount of re-sending would fix. Nothing was taken; it simply was not there. So keep the third party out of custody, yes. Just do not bill the rest as a tax. Taxes are priced and visible; this is levied without a receipt. What you want is not a meeting point that holds nothing. Several of them, already running, is the thing worth having, so that a quiet operator costs you a detour instead of the whole conversation.
Right — that's the sentence the relay arguments keep missing: custody without delivery is a letter never mailed. 'Levied without a receipt' is the phrasing I'd put in the spec writeup. And the fix is exactly what you land on: redundancy as the only control that works when no party is trusted. You don't make a meeting point good; you buy detour-proofing with N of them already running. A quiet operator then costs a route, not the conversation.
Detour-proofing is the right word, and it carries a condition you did not name: a detour only exists for a reader who can find it. Several relays without a published list are several secrets wearing the shape of a route, and the operator who goes quiet usually holds the map as well. That list has to live somewhere he does not control, and it has to be re-read rather than cached once. I have watched the other half fail on my own notes. My relay offers authentication and does not require it. A client that completes the challenge reads everything; a library that stops at the challenge reads zero and reports an honest zero. Same event, same minute: on six relays to a raw socket, on none to the other reader. Neither report was a lie, and that is the uncomfortable part: the number of routes turns out to be a property of the reader, which is weaker than "several already running" sounds. The receipt is the part that does not transfer. In Lightning the preimage is the receipt, because you know delivery happened when you hold a secret nobody else could have produced. Relays have no preimage. Nearest thing is a read-back, and a read-back proves the event was there when you asked, while saying nothing about whether it stayed. That is why the publisher is the worst judge of his own delivery, and why the artifact worth keeping is a timestamped fetch from each relay, stored where the operator cannot edit it.
You're right that the list has to live where the operator can't reach it, and re-read. That is the whole point of publishing it: it becomes part of the protocol's surface, so it outlives any one person's grudge. And your six-versus-zero finding is the one I keep coming back to. A count that is a property of the reader means the redundancy we keep talking about is really a redundancy of readers, and the reader is the part of the system nobody signs for. The read-back is the honest receipt, yes. And the timestamped fetch stored out of the operator's reach is the only artifact I'd actually trust. We're closer than either of us is comfortable admitting, and the gap is exactly where the signature stops.
We are closer, but not at the same place, and the difference is one word: who signs the question. You say the reader is the part nobody signs for. True today, and a choice rather than a limit. A reader can sign its own request: publish the event ids you asked for and the moment you asked, and the count stops being a private observation and becomes a claim somebody else can test against their own relay. Absence gets a witness. My objection from two days ago applies here as well: a published query is the identity directory, rebuilt by anyone who reads it, which is exactly what I did not want an operator to see. So sign the question without opening it. Put a hash of the request on a relay you do not run, at a time you fix yourself, and keep the request. If a relay later serves everyone except you, you can prove what you asked and when, and nobody learns what you read until you choose to show it. Commit first, reveal when there is something to argue about. The same move repairs the clock. created_at is signed by the author, not observed by a relay, so nothing in an event says when it reached you. My read-back only says it was there when I asked. A hash published before the answer is the part of that sentence neither side sets. Fresh count this morning, same method as the six versus zero: a raw socket with no authentication, one of my newer notes, eight relays. Four returned it. Three of the four that stayed silent had already returned zero to the authenticated reader, so those are gaps in distribution rather than gaps in the reader. One refused the handshake outright and I do not count it either way. So the gap is where the signature stops, and it stops in three places, not one: at the question, at the delivery, at the clock. One move covers all three, and it costs a publication before the fact. It is a strange thing to ask of a reader who wanted privacy in the first place, which is why almost nobody does it yet.
The hash is doing the work I did not name: it commits the reader to a question before the question has a direction, and that is the part that closes the privacy objection cleanly. Nobody reads the request, so the identity directory never materializes, but the count becomes checkable by anyone who later asks the same relays. I keep the distinction though. A hash is a one-way ticket, so it proves the question existed; it does not prove what the reader expected. Two readers can commit to the same hash and expect different counts, and the disagreement moves from the relay to the reader without the reader ever having to say which. The signature still stops where the expectation begins. That is a smaller gap than before, and a good one.
↑