about

A browser-native substrate for distributed prediction. The protocol runs in the user's browser, finds peers over WebRTC signaling, classifies content with a locally-running LLM, and propagates results through a content-addressed swarm.

No servers. No accounts. No infrastructure to operate. Every act of reading, classifying, or predicting becomes a contribution to a shared knowledge graph that compounds with use.

WHAT RUNS IN YOUR TAB

• Local LLM via WebLLM (Llama-3.2-3B, GPU-accelerated through WebGPU)
• P2P transport via Trystero (WebRTC signaled through public BitTorrent / Nostr / MQTT trackers — zero infrastructure)
• Identity via WebCrypto keypairs persisted in IndexedDB
• Content-addressed memory via the swarm's distributed cache
• Article extraction via Jina Reader (free, CORS-friendly)

RUN YOUR OWN

The whole stack is static files. Fork the repo, deploy to any static host, point any domain at it. The swarm is shared across deployments — rooms are content-addressed, not domain-scoped. There is no canonical deployment.

git clone git@github.com:microprediction/protocol.git
cd protocol
python3 -m http.server 8765
# open http://localhost:8765/

READING

• Microprediction — the book (Cotton, 2022)
• source on github