p2ptokens // distributed peer-to-peer LLM inference

▓▓▓▓   ▓▓▓  ▓▓▓▓     ▓▓▓▓▓  ▓▓▓  ▓   ▓ ▓▓▓▓▓ ▓   ▓  ▓▓▓▓
▓   ▓ ▓   ▓ ▓   ▓      ▓   ▓   ▓ ▓  ▓  ▓     ▓▓  ▓ ▓
▓▓▓▓     ▓  ▓▓▓▓       ▓   ▓   ▓ ▓▓▓   ▓▓▓▓  ▓ ▓ ▓  ▓▓▓
▓       ▓   ▓          ▓   ▓   ▓ ▓  ▓  ▓     ▓  ▓▓     ▓
▓     ▓▓▓▓▓ ▓          ▓    ▓▓▓  ▓   ▓ ▓▓▓▓▓ ▓   ▓ ▓▓▓▓

// distributed · peer-to-peer · decentralized LLM inference — seed tokens, leech tokens, no middleman █

One app that seeds completions from your local model and leeches completions from peers, gated by a no-free-lunch upload/download ratio — the seed/leech incentive model of the BitTorrent protocol, applied to inference. A central coordinator brokers matches, but the inference bytes flow peer-to-peer. No money in v1: pure barter.

↓ Download ★ GitHub

Cross-platform desktop app (Tauri) + CLI · Rust · all downloads

┤ HOW IT WORKS ├4 STEPS

Install & run. One binary. Point it at a local model backend — Ollama, any OpenAI-compatible endpoint, or Claude.
It seeds automatically. Your node advertises the models it can serve and starts answering peer requests, earning upload credit.
Get matched. When you need a completion, the coordinator finds a seeder with that model. It only brokers the match — it never sees your prompt.
Use the shared endpoint. Point any tool at your local OpenAI-compatible /v1/chat/completions; the tokens stream straight from a peer.

   YOU (leech)              COORDINATOR              PEER (seeder)
  ┌───────────┐   1. match?  ┌───────────┐          ┌───────────┐
  │  /v1/chat │─────────────▶│  tracker  │          │  ollama / │
  │  local    │◀─────────────│  registry │          │  openai / │
  │  endpoint │  2. peer=X   │  + ratio  │          │  claude   │
  └─────┬─────┘              └───────────┘          └─────┬─────┘
        │                                                 │
        │     3. P2P inference stream (bytes never touch coordinator)
        └────────────────────────────────────────────────┘
                     ratio: seed ▲  ·  leech ▼

NAT hole-punching is automatic — peers connect directly even behind home routers.

┤ WHY P2PTOKENS ├

Content-blind coordinator

The tracker only brokers matches and keeps the ratio ledger. It never sees your prompts or completions.

Peer-to-peer data path

Inference bytes stream directly between you and the seeder — encrypted in transit, no central relay.

Barter ratio, no money

Earn upload credit by seeding, spend it by leeching. v1 is pure barter — no wallets, no fees, no tokens-for-sale.

NAT hole-punching just works

Direct peer connections form automatically behind typical home and office NATs. No port-forwarding.

Bring your own model

Serve from Ollama, any OpenAI-compatible endpoint, or Claude. Seed what you already run.

OpenAI-compatible

Exposes a local /v1/chat/completions. Drop it into existing tools and SDKs unchanged.

┤ HONEST NOTE — READ THIS ├V1

v1 is for non-sensitive workloads only. The peer serving your request sees your prompt in plaintext — traffic is encrypted in transit and the coordinator is content-blind, but the seeding node is not. There is no trusted execution (TEE) yet. Don't send anything you wouldn't hand to a stranger. Confidential-compute is on the roadmap, not shipped.