projectpaie/loomprotocol

Loom Protocol is a local first, peer to peer system for breaking a large text processing job into small pieces and spreading that work across ordinary computers on a network, rather than relying on one big server or cloud service. It follows a hub and spoke layout: one coordinator computer, called the hub, splits a document into shards and puts them in a queue, while other computers on the network, called spokes, pull shards from that queue, run them through a local language model, and send back the finished piece. For each shard, two separate local models work in sequence rather than just one. The first model produces a draft, for example an English translation of a passage. The second model reviews that draft for accuracy, tone, and any missing nuance, and a final step folds the review into a corrected version, which then gets recorded along with which machine and which model produced it. Because each spoke only ever receives one small, disconnected piece of a larger document, no single computer or outside model provider sees enough of the whole text to reconstruct it. The README demonstrates the system with a real test: translating a full classic Japanese novel using one Mac Studio and two Windows PCs with graphics cards, running open local models, which took a little under two hours from start to finish and produced tens of thousands of English words. The project is aimed at people who want to process large amounts of text using hardware they already own instead of paying for cloud computing, and it works over local Wi-Fi or a private VPN like Tailscale. It ships as two Python scripts, one to start the hub server and one to start a worker. It is released under the AGPLv3 license, which requires any hosted service built on it to display a notice telling users the software is free and open source.