AI that no single server ever sees whole.
Nepheral is a peer-to-peer mesh where every install is a full node: local inference, a QUIC transport, a lightweight chain, and a privacy layer that splits every query into Shamir shares across independent peers. A deterministic lottery picks who answers — no single relay ever holds a complete message, and no central server runs the model.
- Shamir threshold by default
- 3-of-5
- Shamir threshold by default
- parallel review/research/doc/dev modes
- 4
- parallel review/research/doc/dev modes
- validator today — beta, not decentralized yet
- 1
- validator today — beta, not decentralized yet
01Why
Hosted inference has three problems, not one.
No privacy
Every prompt and reply is logged by the operator. “We don’t train on your data” is a policy, not a property.
No price discipline
Margins on hosted inference run 80–95%. You pay for hyperscaler capacity you never use.
No participation
Compute is dark — you are a buyer, never an operator. There is no path to contribute and earn.
Nepheral is not a replacement for ChatGPT or Claude — it is what you reach for when the threat is the platform operator itself, not a network observer.
02How a query flows
Your prompt, in five steps.
No intermediate party — relays, custodians, or the team’s own authority — ever holds a complete, readable copy of your message.
Encrypt
Your client derives a session key and encrypts your prompt with AEAD (XChaCha20-Poly1305).
Split
The encrypted bundle is split into Shamir shares — 3-of-5 by default — over independent peers. No single share reveals anything.
Draw
A deterministic lottery, seeded by message length and the hour, picks which AI-capable peer may respond. Every custodian recomputes the same result independently before releasing its share.
Prove
The winner must prove cryptographic possession of the winning key before any custodian releases a fragment to it.
Answer
The winner reconstructs your prompt, runs inference locally, and encrypts the reply back to you — and only to you.
The draw is deterministic: every custodian independently recomputes the same result and verifies the winner cryptographically before releasing anything.
03Beyond chat
Four ways to fan a request across the mesh.
Each mode splits one request into several independent sub-requests, dispatches them to different peers in parallel, and reassembles the result. A 4-lens security review finishes in the time of the slowest lens, not the sum of all four.
Review
3–4 lensesSplits one artifact into independent perspectives — auth, injection, secrets, crypto — each reviewed by a different peer.
Doc
up to 8 chunksSplits a long document by paragraph and processes chunks in parallel — summarize, extract, or search across all of them at once.
Research
3–4 anglesSends the same open question to peers primed with deliberately opposing angles, then surfaces where they agree and where they do not.
Dev
up to 12 filesSplits a multi-file task by file, one worker per file, with an explicit instruction not to reference the others.
04Under the hood
One binary, three roles.
Every install is simultaneously a client, a relay, and — if it can run a model — an inference responder.
Transport
QUIC over UDP/443, so it passes through networks that only allow HTTP/3-shaped traffic. IPv4 and IPv6, mDNS on desktop, and a relay for peers behind NAT.
Privacy
AEAD encryption, per-recipient key wrapping over X25519, and Shamir splitting — before anything leaves your machine.
Settlement
A lightweight chain records contribution and consumption per peer and mints credits at a calibrated, adjustable rate.
Drop-in for your existing tools
A local daemon on 127.0.0.1:11434 — the same port Ollama uses — speaks both the OpenAI and Ollama APIs. Point Continue.dev, Aider, Cursor, or Cline at it and it just works.
127.0.0.1:1143405Economy
Credits, not subscriptions.
An on-chain ledger tracks what you contribute and what you consume, at rates that are parameters — not promises.
- 1,000,000
- N-AI genesis treasury
- 10 credits
- = 10 inferences on a top-tier session, for 1 hour
- 0
- transfer fees — the chain charges no gas
- ~7 days
- timelock before a price change takes effect, visible on-chain in advance
Supply is elastic by design: emission targets a stable credits-per-hour rate and adjusts itself from real network activity.
06Shipped
What actually landed, in plain language.
Distilled from real commits — no vanity metrics, no marketing filler.
One binary for any GPU
The build now detects NVIDIA, AMD, or CPU-only at startup and loads the right backend — no more choosing the wrong download.
The mesh reaches peers behind NAT
A relay node can now forward an encrypted envelope it can never open, so a lottery winner behind NAT is still reachable.
Mesh chat goes from broken to usable
Five root causes fixed in one pass: capability gossip was too slow, half the discovered connections were write-only and silently dropped replies, and the share pool counted dead connections.
Four modes no single model can imitate
Review, Doc, Research, and Dev ship: splitting one request into parallel, independently-answered pieces across the mesh.
The lottery stops being predictable
The original draw was a public hash anyone could precompute. A verifiable-key proof now gates every claim.
Real economics instead of fixed constants
Genesis treasury, an authority pallet, per-win weighting, and a self-adjusting conversion rate replace hardcoded numbers.
07Right now
We are recruiting testers and operators.
This is beta software, pre-audit, built by a small team moving fast. If that is a feature to you and not a bug, we want you in the mesh.
Run a node
Download the desktop client or the headless node, join the mesh, and tell us what breaks.
Run a top-tier channel
Have a spare GPU — an H100 or similar — and want to serve the large-model channels? We are looking for the first attested operators.
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.