raghavan/agent-torrent — explained in plain English
Analysis updated 2026-05-18
Delegate a coding task from one machine to another peer running Claude or Codex.
Share idle AI coding agent capacity with peers and earn credits in return.
Run coding tasks against a locally hosted language model with no cloud account.
| raghavan/agent-torrent | 1038lab/agnes-ai | 3eyedtiger/video2vrcemote | |
|---|---|---|---|
| Stars | 4 | 4 | 4 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | easy |
| Complexity | 4/5 | 2/5 | 1/5 |
| Audience | developer | vibe coder | vibe coder |
Figures from each repo's GitHub metadata at analysis time.
No authorization or encryption layer yet, the author recommends trusted networks or a VPN only.
Agent Torrent is a research prototype that lets computers running AI coding agents, like Claude or Codex, share their idle capacity with each other over a peer to peer network, similar in spirit to how BitTorrent lets computers share files directly instead of through a central server. The idea is that most people's AI coding subscriptions sit unused for much of the day, so this project lets machines offer that spare capacity to other peers on the network and earn credits they can later spend when they need extra capacity themselves. There is no central server and no single coordinator. Every computer running the software, called a peer, can act as both a requester asking for help and a worker doing the work, using the same code. Peers find each other by broadcasting signals on the local network every few seconds, and can also be pointed at specific peers across the internet. Every message peers send to each other is signed using cryptographic keys, so a peer can verify who sent a message before acting on it. When one peer delegates a coding task to another, it puts up one credit as a kind of deposit, which gets released to the peer that completes the work successfully, or returned if the task fails or times out. The task itself runs inside a sandboxed, temporary environment with a strict time limit, so the worker's system stays isolated from whatever it is asked to run. The project explicitly states it is an early research prototype: there is currently no authorization system controlling who can send tasks, no verification that results are correct, and no encryption on network traffic. The author's documentation is direct about this and recommends only running it inside a trusted network, or over a private VPN connection, rather than exposing it to the open internet. It supports the Claude and Codex command line tools, as well as any locally run language model server, so it can work without needing a cloud account at all.
A peer to peer network letting idle Claude or Codex coding agent capacity be shared between computers, who earn credits in return.
Mainly Python. The stack also includes Python, PyNaCl, Ed25519.
MIT license: free to use, modify, and distribute, including commercially, as long as the copyright notice is kept.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.