AI no server sees whole.
Your question is encrypted on your machine, split into fragments held by independent nodes, and answered by a model running on someone else's machine. Nobody along the way ever holds the whole message.
- node holds fragments of every question
- Every
- node holds fragments of every question
- servers that see your whole message
- 0
- servers that see your whole message
- desktop, server and phone (alpha) · macOS soon
- Win · Linux · Android
- desktop, server and phone (alpha) · macOS soon
βDownload
Get the beta.
Pick your system and, on a computer, how the model should run. Every install is a node: it joins the mesh, serves it, and keeps your wallet behind your own password. A phone runs no model: it holds fragments and asks the network.
3 · Your download
On first launch
You create a wallet: 24 words shown once, and a password. Copy the words by hand on paper — they are the only way back in. Nobody from Nepheral will ever ask for them.
Beta, on the test network
Pre-audit software on the testnet. Builds are not code-signed yet, so Windows SmartScreen warns: More info → Run anyway.
Source code
The code of the desktop app and the node is public: read it, audit it or build it yourself. BUSL-1.1 today; each version becomes Apache 2.0 four years after its release.
github.com/JEHR037/Nepheral-core →
01Why
Hosted AI has three problems, not one.
Nepheral does not replace ChatGPT or Claude: it is what you reach for when the adversary is the platform operator itself, not an observer on the network.
- 01
No privacy
Every prompt and every answer is logged by the operator. "We don't train on your data" is a policy, not a property.
- 02
No price discipline
Hosted inference margins sit around 80–95%. You pay for hyperscaler capacity you never use.
- 03
No participation
Compute is opaque: you can only buy it, never run it. There is no way to contribute and earn.
02How a question travels
Your message, in five steps.
No intermediary — relays, custodians, or the team's own nodes — ever holds a complete, readable copy of your message.
Behind home routers, nodes connect directly by hole punching. When a router will not allow it, fragments travel sealed through public nodes that cannot open them.
01
Encrypt
Your machine encrypts the question with AEAD (XChaCha20-Poly1305) for the nodes that may answer it, after swapping the personal data it recognises for placeholders.
02
Split
The encrypted bundle is split into Shamir fragments across every node on the network, with a model or without. Rebuilding it takes at least half of them, and none alone reveals anything.
03
Draw
A reputation-weighted draw picks which AI node answers. Every custodian recomputes it on its own before releasing its fragment.
04
Prove
The winner proves with a VRF proof that the draw picked it. Without that proof, nobody hands it a fragment.
05
Answer
The winner rebuilds the question, runs the model on its own machine, and sends the answer back encrypted to you alone.
03Inside
One binary, three roles.
Every install is a client, a custodian and — when it can run a model — a node that answers.
- QUIC
- Encryption
- Shamir
- VRF draw
- Chain
- Local API
Transport
QUIC over UDP with mutual TLS bound to each identity. IPv4 and IPv6, hole punching between peers behind NAT, and sealed relays when it cannot be done.
Privacy
AEAD encryption, per-recipient key wrapping over X25519 and Shamir splitting, before anything leaves your machine.
Settlement
A lightweight chain records every answer, signed by the node that gave it and the one that asked, and mints credits at a calibrated rate.
Works with the tools you already use
A local daemon on 127.0.0.1:11434 — the port Ollama uses — speaks both the OpenAI and the Ollama API. Point Continue.dev, Aider, Cursor or Cline at it and it just works.
127.0.0.1:1143404Economy
Credits, not subscriptions.
An on-chain ledger records what you give and what you use, at rates that are parameters, not promises.
- 1,000,000
- genesis treasury in N-AI
- 10 credits
- = 10 inferences in a top-tier session, for 1 hour
- 0
- transfer fees: the chain charges no gas
- ~7 days
- before a price change takes effect, visible on-chain in advance
Supply is elastic by design: emission tracks a steady rate of credits per hour and adjusts itself to the network's real activity.
05Right now
We are looking for the mesh's first people.
This is pre-audit beta software, built by a small team that moves fast. If that sounds like a feature rather than a bug, there is a place for you.
Testers
Install the desktop app on Windows or Linux, ask it things, break it, and tell us what happened.
With a graphics card
If your machine has an NVIDIA or AMD card, it answers other people's questions much faster and earns credits with each one.
A server or VPS
The headless node runs as a service on Debian or Ubuntu and keeps the mesh up while you are away.
From your phone · alpha
The Android app holds fragments and lets you ask the network, without running the model on the phone.
Where things stand: Nepheral is beta and not yet audited, and the validator set starts centralized while the network grows. That is what you would be joining early to help change.
06FAQ
What everyone asks.
What do I need to run a node?
Windows 10 or 11, or 64-bit Linux (Ubuntu 22.04+, Debian 12+), and 8 GB of RAM. A graphics card is optional: with NVIDIA or AMD, answers come much faster.
Will there be a phone app?
Yes, for Android, in alpha: it is under Downloads. It does not run the model on the phone: it holds fragments of other people's questions, earning credits for it, and lets you ask the network from anywhere. iPhone later.
Can anyone read my conversations?
No relay or custodian can: they only see encrypted fragments. The node that answers does read the question, because it has to run the model, but it only sees an anonymous node key, not who you are. Personal data the app recognises is swapped for placeholders before it leaves your machine, and conversations are not logged.
What do I earn by keeping a node on?
Every answer your node gives is signed by the node that asked and credits you N-AI on the chain. Without a model, your node holds fragments and makes the network sturdier.
Does it work behind my home router?
Yes. Nodes connect to each other by hole punching through the public nodes. If your router will not allow it, fragments reach you sealed through a relay that cannot open them.
Why does Windows warn me when I open it?
Beta builds are not code-signed yet, so SmartScreen warns: choose "More info", then "Run anyway". Signing is on its way.
Can I use a machine with several graphics cards?
Yes: today a node uses all your NVIDIA cards to load a larger model. Answering a different question on each card at once is next.
Does it work with my tools?
The desktop app opens a local API on 127.0.0.1:11434, compatible with OpenAI and Ollama. Continue.dev, Aider, Cursor or Cline work pointed there.
The mesh grows with every node.
Install it, leave your node on, and answer others while you sleep.