An operating system whose identity and configuration live on Nostr relays rather than on the device: what already runs, what it would actually take — bootloader architecture, carrier certification, unfunded drivers, and an EU that named the lock but legislated around it — and why twelve words in your head are more yours than any handset.
There is a project that boots a computer by asking for twelve words. Type them and the machine assembles itself — files, contacts, settings, identity. Power it off and it forgets everything. Its root filesystem lives in memory. Nothing is written to disk, not even the Wi-Fi password.
It runs on laptops and Raspberry Pis. It does not run on phones. The distance between those two facts is what interests me here.
It is called n-OS-tr. Those twelve words are a BIP-39 mnemonic, and from it the system derives every key it needs along the standard NIP-06 path — one index for the user's identity, another for the admin key of a local relay, a third for a blob store, a fourth for a mesh node. Preferences and dotfiles live on Nostr relays as replaceable events, encrypted to yourself under NIP-44, and get projected into the home directory at boot. As of today the identity layer is still in design; the mesh daemon, the relay and the blob server are running.
What makes it interesting is not the ephemerality. It is the direction the state travels.
A normal operating system is its state. Your contacts are a database on your phone. Your photos are a directory. Your identity, where it exists at all, is an account whose master copy somebody else holds. Reinstall the system and you have destroyed the machine; restore it from a backup and you have borrowed one.
In the inverted model the device holds nothing that defines you. The definition sits elsewhere, signed, and the device is a cache that fills at boot and empties at shutdown. That reframes what a phone is. It stops being a thing that accumulates your life and becomes a thing that renders it.
Numbers make this less abstract than it sounds. I measured a follow list off the relay network myself to see what a piece of a life costs. Measured 28 September 2026: 695 contacts in one event, 49.9 kilobytes, roughly 74 bytes per person. Same run: a profile carried 675 bytes, an average note about one kilobyte, a long-form article around ten. A timeline of five hundred notes — an afternoon of scrolling — came to 591 kilobytes. An entire social graph, the asset platforms treat as their crown jewel, is smaller than one photograph.
That asymmetry carries the argument. That state — the part that makes you you, in the sense a platform cares about, is tiny, cheap and portable. What is genuinely large — photographs, video, the actual megabytes — is also the state nobody needs to hold on your behalf.
None of this requires a new operating system to be worth talking about today. The stack is real and it runs on hardware you can buy.
GrapheneOS is a hardened Android that strips Google out of the base and sandboxes Play Services when an app insists on them. It runs only on Pixel devices, and the reason is not laziness: Pixels are the only phones with unlockable bootloaders that still support proper verified boot after you flash something else. The irony — the most Google-free handset is a Google phone — is a hardware constraint, not a preference. Motorola announced a partnership at MWC 2026, with first models expected in 2027.
On top of that system, Amber holds the private key in one place and signs for other apps through NIP-55, so a compromised client can lie to you but cannot steal your identity. Zapstore publishes applications as Nostr events under NIP-82: the developer signs the release, the client checks that signature against a public key before installing, and discovery comes from the people you follow instead of from a ranking system. Citrine runs a relay on the phone itself, so drafts and bookmarks have a home that answers to you.
That is a coherent stack. It is also still an Android phone with a Google-derived kernel, Google's permission model, and Google's release cadence. In 2026 Google moved AOSP source publication from quarterly to twice a year, Q2 and Q4. That upstream is narrowing while everyone builds on it.
Here the idea meets something harder than software.
A phone is not a small computer. It is a radio with a computer attached, and the radio is the proprietary part. This baseband processor — the chip that talks to the cell network — runs closed firmware that you cannot audit, cannot replace, and on most devices cannot even observe. It has direct memory access. It has been the subject of exploitation research for two decades. Your operating system can be perfectly sovereign and still share a bus with a component that answers to nobody you can name.
This is why the honest description of a de-Googled phone is reduced dependency, not eliminated dependency. Its one realistic path to a genuinely open modem is the Quectel EG25-G inside the PinePhone and PinePhone Pro, and that path is narrow. Modem firmware updates took years to become routine. Camera support is degraded because most mobile image signal processors have no mainline driver at all. People building the Linux phone ecosystem describe their own devices as development targets that enthusiasts run daily — real, decade-long work, funded thin, with no device supported perfectly.
So the layer where a phone is most exposed is the layer where the Nostr idea has the least to say. Signing events beautifully does not help you if the component transmitting them is a black box.
The second hard part is subtler, and I have not seen it solved anywhere.
A phone's value is that it interrupts you. That is not a side effect of the product; it is the product. Interruption needs standing infrastructure — something must be watching on your behalf while your device sleeps, so that the radio can stay off and the battery can survive until evening.
On Android that watcher is Firebase Cloud Messaging. One party, chosen by the system vendor, learns when every app wants to reach you and roughly where you are. The de-Googled answer is UnifiedPush, and it works: ntfy can act as a distributor, it is self-hostable, it can run entirely inside your own network.
Now look at what you have done. You did not remove the watcher. You became one. Your own server sees the metadata Google saw — which app is trying to reach you, how often, and when you are awake. Strictly better if you are the operator. Most people are not. The alternative is a persistent connection, which means the modem is always on, which is precisely what the push system existed to avoid.
There is no third option on the table. Interruption requires an intermediary, and the intermediary learns the shape of your attention. On iOS you do not even get the choice: Apple's push service is mandatory, which is why the same de-Googled stack cannot be built there at all.
Naming the obstacles is cheap. The interesting question is what the things being missing actually are, and which of them are missing for technical reasons rather than commercial ones.
Start with the bootloader, because that is where the whole idea either lives or dies. An unlocked bootloader disables verified boot, and verified boot is what lets a bank app trust the device it is running on. Open the bootloader and the banking apps, the DRM, the government ID — all of it stops working, by policy. That is why most manufacturers lock it and why the lock is not simply corporate malice: a phone that cannot prove what it is running cannot be trusted with the things people keep on phones.
Fairphone found a way through, and the design is worth naming precisely. They isolate the radio firmware into a protected section, verified and locked entirely separately from the Android system. Because a custom ROM never touches the isolated modem code, the manufacturer can keep the bootloader unlockable without the radio certification falling apart. That is the trick: separate the part that must stay closed from the part that does not need to be, and audit them differently. It is not a software feature. It is an architecture decision, made years before a customer sees the phone.
The second thing that would be needed is carriers willing to allow a device that is not Android. This is where the practical ceiling shows. A PinePhone running Ubuntu Touch may lack VoLTE while the same hardware on Sailfish OS gains it — because Jolla licenses proprietary modem firmware. In the United States it goes further: AT&T blocks non-certified IMEI ranges outright, and T-Mobile and Verizon want IMS profiles and eSIM provisioning that are rarely implemented outside Android. A phone that cannot place a call is not a phone, no matter how well it signs events.
The third is money, and it is the least glamorous. postmarketOS has no device supported perfectly. Most mobile image signal processors have no mainline driver at all, so the camera is either missing or reduced to a still-image capture — on more than four in five supported phones by the project's own account. That is not a hard problem in the sense of being unsolvable. It is a hard problem in the sense of being unfunded: the maintainers describe their own work as dependent on free time, and they can name what targeted funding would buy.
And then there is the law, where the finding surprised me.
The EU wrote the problem down. Recital 7 of its smartphone ecodesign regulation states plainly that "it is currently not possible, or extremely difficult, for the owners of mobile phones, including smartphones, and tablets to change the operating system of their device, which is chosen and maintained by the manufacturer". A regulator that names the lock is a regulator that could open it.
It did not. The regulation mandates five years of operating system upgrades, spare parts availability for seven years, battery endurance thresholds, and a ban on parts pairing — all of which make a phone last longer. On the bootloader, the word appears exactly once in the entire text, and only as one possible place to implement secure erasure of encryption keys. The right to repair directive, applicable across the EU since 31 July 2026, goes further on parts and pricing. Neither creates a right to install another operating system, and neither requires a manufacturer to keep the bootloader unlockable.
Read the recital next to the operative articles and the shape is clear. The EU decided that phones should last longer, and that the choice of operating system was not its business. It legislated durability and left sovereignty alone. That is a coherent position. It is also the reason the sovereign phone is currently an enthusiast project in Europe and a hardware experiment in China, rather than a product category with a legal floor under it.
What would change it is narrow and specific: a requirement that the bootloader remain unlockable where a manufacturer's radio isolation already makes it possible, and that devices which pass certification not be refused by carriers for lacking an Android lineage. Both are regulatory, not technical. Fairphone has already proved the engineering half. Nobody has written the other half down.
The identity half has its own missing piece, and it is not cryptographic. Twelve words in a head are one lost memory away from gone. Steel plate solves storage but not inheritance, and inheritance is where every self-custody design quietly stops. The workable answer already exists in the Bitcoin world and has never been ported into an operating system: threshold recovery, three shares with two needed, held by people who know what they are holding and survive you. That is a social arrangement wearing a cryptographic coat. It needs no new protocol — NIP-06 already derives the keys. It needs a default that a person who is not an engineer will actually follow on the day they set the phone up.
Two things I could not verify, and I would rather say so than round them off.
One is that a phone which forgets you is a good phone. It is elegant until you are standing in an airport on a flat battery with a mnemonic in your head. "Carry it in your head" does not survive a car crash. The answer from that project is that identity is portable, which is true, and that the device is disposable, which is also true. Neither fact addresses the thing you have actually done: moved your single point of failure from a device you can insure to a memory you cannot back up.
The other is the timeline. Every part of the sovereign stack that depends on hardware is years out, gated by chip vendors with no commercial incentive to open their modems. Every part that depends on Android is exposed to a narrowing upstream. A confident five-year prediction would be dishonest. What I can say is that the software half is much further along than most people assume — identity, signing, distribution are built and running — and that the hardware half is barely moving.
Whether a Nostr-based phone can be built is not the interesting question. Much of it has been. What a phone is for, once you take the idea seriously, is.
A phone is the device that is always with you, always on, always reachable. That is also a precise description of the device that leaks the most about you. No amount of cryptographic identity changes the arithmetic: reachability needs a listener, the listener learns something, and the listener is not you.
Maybe the conclusion is that a phone is the wrong place to insist on sovereignty — that the sovereign device is the one you leave behind, and the thing you actually carry is not a machine at all. Twelve words stamped on steel, or held in memory, are more yours than any handset and they fit in a pocket. Everything else is a cache: warm, useful, and disposable.
Which is the same claim n-OS-tr makes. It just makes it on a laptop, where the radio cannot argue.
