ProjectZero

Pool the spare computers in your home, club or community into one small AI grid. Chat with an AI that runs on your group's machines, earn credits by sharing your own computer, and talk to each other in a lobby, in groups or directly.

Deutsch · Contributing · Security · Roadmap · Vision · MIT license

Status: experimental (v0.3). It works on a home network today and is tested on Linux, macOS and Windows. It is not a hardened public service. Requests are readable by the machine that computes them and by whoever runs the server. Credits are play money, not money. See what it is not.

What you get

• AI chat that spends credits. Every answer costs a few credits and is computed by one of the machines that are online right now. You see which models are available.
• Credits for sharing your computer. Switch on "Share computing power" and your PC answers other people's requests while you are not using it. Your balance is always visible on the right.
• You decide how much you share. Sliders for CPU threads, GPU layers, answer length and context size. Changes apply immediately.
• Lobby, groups and direct messages with everyone on the same server, plus a list of the machines and people online.
• English and German, light and anthracite theme, works on a phone-sized screen.
• Self-hosting in one command. Run your own server for your family, flat share or club.

Install

You need Python 3.11 or newer. The installer creates its own environment in your user folder, needs no administrator rights and changes nothing system-wide.

Linux and macOS

curl -fsSL https://raw.githubusercontent.com/maun200/ProjectZero/main/install.sh | sh

Windows (PowerShell)

irm https://raw.githubusercontent.com/maun200/ProjectZero/main/install.ps1 | iex

The app opens in your browser. Start it again later with projectzero. Prefer to read a script before running it? Download install.sh / install.ps1 first, or install manually with pip install https://github.com/maun200/ProjectZero/archive/refs/heads/main.tar.gz.

Try it in two minutes, without any AI model

projectzero demo

This starts a private server and the app on your computer only. Create an account, switch on "Share computing power" with execution "Demo", then ask the AI something. The answer is a placeholder, but everything else is real: queueing, your machine picking up the job, credits.

Join a server and share real AI

1. Ask whoever runs your group's server for its address, for example http://192.168.1.10:8765 on your home network, or an https:// address on the internet.
2. Start projectzero, enter the address, choose a display name.
3. To share your computer, install Ollama, pick one of the models the server allows and switch on "Share computing power". The app can download the model for you.

Run your own server

projectzero server --lan --models qwen2.5-coder:1.5b

The server stores its keys in your user folder (readable only by you) and allows self-registration by default. Useful options:

Option	Meaning
--lan	reachable from your local network (default: this computer only)
--registration open|invite|off	who can create an account; invite needs codes from POST /api/admin/invites
--models a,b	the models members may use and share
--auto-settle	credit finished jobs to the machine owner immediately (nobody checks the answers!)
--signup-credits N	credits a new member starts with

For an always-on server there is a hardened Docker setup:

cp .env.example .env    # then set two random keys and CAG_REGISTRATION / CAG_ALLOWED_MODELS
docker compose up -d --build

Do not put the server on the internet without an HTTPS reverse proxy and reading docs/SECURITY.md.

How it works

 your browser ── local app (127.0.0.1) ──┐
                                         │  HTTPS or private network
 other members' apps ───────────────────►  coordinator (FastAPI + SQLite)
                                         │   queue, accounts, credits, chat
 machines sharing compute ◄──── pull ────┘
   └─ local model via Ollama (never exposed to the network)

• Workers pull jobs; the server never connects into your computer and never sends code to run. A worker only feeds the request text to the local model you picked.
• Each job goes to one machine that has the whole model. Splitting one big model across many machines is research, not a feature (see docs/RESEARCH_GLOBAL_GRID.md).
• Leases, heartbeats, retries and refunds keep jobs moving when a machine disappears.
• privacy=trusted jobs only go to machines the operator marked as trusted; everyone who registers through the app is a community machine.

Where this is going: a global community grid

ProjectZero starts small on purpose: one server for a household, a club or a group of friends. The long-term goal is bigger: a global, community-owned AI network that is a real alternative to commercial AI providers. People everywhere contribute hardware they already own, and the network does not depend on any single company or server.

The plan is federation, similar to e-mail or Matrix: many independent servers, each run by its own community, that forward work to each other when they are busy and settle credits between them. A single answer will not be as fast as from a data center. In return the network can be almost free to use (the main cost is the electricity of hardware that already exists), independent and open.

None of this exists yet. Today every server stands on its own. The steps to get there are Phase 3 of the roadmap and are worked out in docs/RESEARCH_GLOBAL_GRID.md. If you want to help build it, that is the place to start.

What it is not

• Not private. Prompts and chats are stored unencrypted on the server and the computing machine can read the request. Do not send personal or secret data.
• Not verified. Nobody checks that an answer was computed honestly. With --auto-settle credits are paid without any check.
• Not money. Credits only decide who may use how much compute. They cannot be bought, sold or transferred.
• Not a public service yet. There is no abuse protection beyond rate limits, invite codes and an admin block function.

Development

git clone https://github.com/maun200/ProjectZero && cd ProjectZero
python -m venv .venv && . .venv/bin/activate
pip install -e '.[test]'
pytest -q

Project layout: coordinator/ (server), worker/ (job runner), projectzero/ (local app and CLI), gui/ (optional tray app for headless workers), scripts/ (admin and device helpers), tests/. More in CONTRIBUTING.md and docs/ARCHITECTURE.md.

Bug reports, device reports (which hardware runs which model how fast) and pull requests are very welcome.

License

MIT, created by maun200. You may use, change and share it as long as the copyright notice stays with it. See LICENSE.