martin123132/chatp2p

ChatP2P is an early prototype of a peer-to-peer network where regular computers contribute AI compute power to each other. The project focuses on the unglamorous plumbing that has to work before any larger vision is possible: signed nodes proving who they are, signed jobs proving what work was asked for, verified results, a credit system, and a central coordinator that may later give way to a wider mesh of peers. In its current form, a coordinator process creates signed job packets and a worker node signs its registration. The worker verifies any job it is leased before running it, then signs the completed result. The coordinator then verifies that result and awards credits. Network liveness and job leasing also use signed packets, so every step has a cryptographic trail. Users install the package with pip and run a command line tool called chatp2p. There are commands to benchmark a machine before it joins, including hardware checks, GPU detection, and a list of locally installed Ollama models. A doctor command prints JSON checks for identity, the local benchmark, Ollama reachability, and coordinator reachability. The coordinator keeps state in a small SQLite file, so registered nodes, jobs, leases, results, and credits survive restarts. A local web dashboard at port 8765 and a handful of JSON endpoints expose the current state. The project ships several built-in job types: simple arithmetic, number theory, text tasks, plain echo prompts, and inference jobs that call a local Ollama model such as llama3.2:3b. Inference jobs are only leased to workers that have advertised the matching model, so the network does not send work to machines that cannot run it. A reliability proof tool can spin up many workers on one machine and run hundreds of jobs to test that the lease and verification flow holds up, including faulty workers that vanish mid-job. A separate public alpha mode adds an admission token, a private operator config, and an invite file so a coordinator can be exposed beyond one machine without becoming open to the internet. The README also covers how to use Tailscale for the first remote partner test.