Lending, in full

How lending works

woodls host runs an open model on your computer, in woodls’ own engine, inside a sandbox that can’t open your files or reach the internet by itself. It serves only while you’re away: after five minutes without your keyboard and mouse, and on a laptop only on mains power. It steps aside the moment you’re back, when another app needs the graphics chip, or when memory runs short, and starts again after five more quiet minutes.

On the public network, each job your computer runs earns credits. In your own private groups it serves your computers and your friends’, for free. What lending does to your computer has more, with an electricity calculator.

Which model fits your computer

woodls models

Every model woodls offers, its download size, the memory it needs, whether a computer serves it now, its price, and whether it fits this computer.

woodls gives models two thirds of a Mac’s memory, keeping the rest for you. A model fits when its file and the room it needs to work fit in that. --memory 30GB gives models more, up to what macOS allows; a 36 GB Mac lends Qwen3.6 35B only that way.

On a PC, Qwen3 8B runs on an NVIDIA graphics card with 8 GB or more (with 8 GB, its window is 8,192 tokens), on a recent AMD card under Linux itself, or more slowly on the processor.

Your first start

woodls host

The first time, with no model named, woodls host:

1. Looks for a model that fits in your Hugging Face cache (~/.cache/huggingface/hub), where other tools may have put one already. It finds a file by its fingerprint, whatever it’s called.
2. If there’s none, offers the one that fits, saying its size, where it goes and the room there, and downloads it after your yes. Every byte is checked against woodls’ SHA-256.
3. Starts woodls’ engine with it and asks woodls’ server to admit this computer: five short prompts whose answers must match ones recorded on a Mac, at 20 tokens a second or more.
4. Once admitted, lists it on the public network and shows the host view.

A plain woodls host starts the same way again next time. To choose the model yourself:

woodls models get

Download the public model that fits this computer

woodls models get qwen3-8b

Download one by name, or any part of it

woodls host --gguf FILE

Lend a model file you already have

A download that stops carries on from where it stopped the next time you run it. It uses Hugging Face’s own cache layout, so llama.cpp, Hugging Face’s tools and woodls code’s --this-computer find the same file, and nothing is downloaded twice.

The host view

In a terminal, woodls host runs in the background and shows the host view: what it serves and to whom, why it’s paused or asleep, its jobs today, what it has earned, and whether woodls’ server lists it now (“listed for gpt-oss 20B · 2 computers serve it now”).

p

Pause for an hour, the model kept loaded

r

Resume

s, then s again

Stop lending

esc

Leave the view; it keeps serving

What its states mean:

When it serves

woodls pause --for 2h

Pause for a while, or until you resume; the model stays loaded

woodls resume

Lift your pause

woodls host --when 23:00-08:00

Serve only in these hours, by this computer’s clock

woodls host --always-on

Keep serving while you use it, never on battery

woodls host --status

What it’s doing, in words

woodls host --stop

Stop lending on this computer

--rejoin-after 600 makes it wait ten minutes after your last touch instead of five. A game played only with a controller isn’t a touch: on Windows, run woodls pause before you play.

Keep it running

woodls host --install

Lends as a background service with the settings you give it: a launchd agent on a Mac, a systemd user unit on Linux, and on Windows a Startup entry that starts lending, hidden, when you sign in. woodls host --uninstall takes the service out.

Sleep. While it may serve, on mains power, woodls asks your computer not to sleep for being idle; the screen still turns off. It lets go on battery, when you pause it and outside your hours. A laptop with its lid closed sleeps all the same. woodls config set prevent-sleep off turns this off.

Notices. woodls config set notify on shows a notice when it earns, when lending stops on a problem, when woodls’ server doesn’t admit it, and when a check of its work fails.

What it earns

Every job your computer runs on the public network earns credits, at the model’s price for the text it read and wrote, less a 15% fee to the shared account. Then:

• Pending first: credits stay pending for 14 days while your computer is new, then for 7, before you can spend them. A computer stays new until 100 of its jobs have passed their check: woodls runs jobs again on its own computers and compares the answers.
• A daily limit: each computer may earn up to 500 credits a UTC day.
• A job whose check fails is paid back to the person who asked. A computer whose checks keep failing is suspended from the public network, and its pending credits are forfeited.
• How much: it depends on how busy the network is, which varies.

woodls status shows what this computer earned today, woodls credits every entry and what’s pending, and Credits and spending the rest.

Lending only to your groups

woodls host --no-public --group home

Serves your private group home only: no admission, no credits, any model file on your own computers. --group may be given more than once. Without --no-public it serves the public network too. A friend’s computer is served only the models woodls has priced: Private groups.

On Windows

woodls host in PowerShell starts lending inside Ubuntu through a hidden helper that Windows holds, so closing PowerShell doesn’t stop it. woodls host --install adds a Startup entry, which needs no administrator. While it lends and the PC is plugged in, woodls keeps Windows from sleeping. It reads Windows’ own keyboard and mouse, power and graphics use.

• Lending from Windows is for your private groups for now.
• woodls uses an NVIDIA card through Windows’ own driver. It fetches its engine for the card the first time, about 1.3 GB, checked against woodls’ signed list.
• Inside WSL2, woodls can’t see a game start until lending has paused: run woodls pause before you play.
• If Windows couldn’t hold lending, the host view says “esc leave: hosting ends with Ubuntu’s last window”: keep that window open, and see troubleshooting.

On Linux

The first time you lend, woodls finds your graphics card and fetches the engine for it, checked against woodls’ signed list: CUDA 12 or 13 for NVIDIA (1.3 to 1.6 GB), ROCm for a recent AMD card (0.34 GB), or the processor’s (59 MB). If a card’s engine stops before the model loads, the processor’s takes over and woodls says why. --engine-variant cpu asks for the processor’s.

Lending needs x86-64 Linux from 2024 on, and the engine runs only in its sandbox. Lending from Linux is for your private groups for now. --no-sandbox runs the engine outside its sandbox: it’s for development only, and woodls doesn’t stop such an engine taking others’ requests, so never lend that way to anyone but yourself.

Stopping and removing

woodls host --stop stops lending now; an installed service starts it again at the next sign-in, until woodls host --uninstall. woodls uninstall stops lending and takes woodls away. Models stay in the Hugging Face cache, where other programs share them: delete them there if you want the room back.