PoolParty — every GPU in the house, one AI

Three clicks, no configuration

Machines find each other on their own. There are no IP addresses to type, no config files, and no separate model downloads per PC.

Install on each PC

One installer, same on every machine. The AI engine ships inside it, so every PC runs an identical build — the usual cause of cross-machine crashes simply can't happen.

Press Share my GPU

That machine announces itself on your network and its graphics memory joins the pool. Integrated graphics are skipped automatically, so the total is honest.

Start the AI server

Pick a model from the list — it downloads itself. When it's ready you get a chat page in any browser on the network, plus endpoints for apps that expect OpenAI or Ollama.

What you actually get

A chat page for the whole house

Open the address on a phone, a tablet, or the PC in the kitchen. Same pool, same conversation history, no software to install on the device.

Code arrives as files

Ask for a program and the files land in your Documents folder — named, foldered, and ready to run. No copying out of a chat window.

Your existing AI apps just work

Anything that speaks the OpenAI API points at the pool. Apps that look for Ollama find it on the usual port without being told.

A model list that fits your hardware

Models too big for your pool are greyed out instead of failing halfway through a download. The best one that fits is picked for you.

Measured on a real house

Three ordinary Windows PCs — a desktop with a 12 GB card, a laptop with 8 GB, a work machine with 7.8 GB — wired to one gigabit switch. 27.8 GB pooled. Every number below came off that setup, not a datasheet.

Model	Size	Reading your prompt	Writing the answer
Llama 3.3 70B	24 GB	79 tok/s	3.2 tok/s
Qwen3.6 35B	20 GB	79 tok/s	5.5 tok/s
Qwen3.8 27B	17 GB	21 tok/s	8.6 tok/s

For comparison: none of these three machines can run any of these models on its own. The largest single card in the house holds 12 GB.

What it costs you

Pooling buys capacity, not speed. If a model already fits on one of your cards, run it there — it will be faster. Here is the honest list.

Single-digit words per second. Every word travels through every machine in turn. Expect a paragraph to take a while and a long program to take several minutes.
Wired beats wireless, by a lot. Wi-Fi works, but latency between machines is the bottleneck. A $20 switch is the best upgrade you can buy for this.
Local models are not frontier models. A 27B model running on household hardware writes useful code and answers well. It is not going to match the best hosted assistants, and pretending otherwise wastes your evening.
Every PC must stay awake. Sharing machines have to be on, unlocked enough to run, and not asleep. Close the app and that card leaves the pool.
Private network only. The engine that shares a GPU has no authentication of its own, so PoolParty refuses to run on networks Windows marks Public and never opens itself to the internet.

Get PoolParty

Download the installer and run it on each PC you want to contribute a card. Windows will warn you that it isn't signed — choose More info → Run anyway. To build it yourself instead, you need Rust, Node.js, and the Visual Studio build tools:

# fetch the AI engine (once)
powershell -ExecutionPolicy Bypass -File scripts\fetch-llamacpp.ps1

# build the installer
npm install
npm run tauri build

# installer lands in
src-tauri\target\release\bundle\nsis\

Releases Source on GitHub