Install the node
One binary that sits beside the inference server you already run. It does not replace Ollama or vLLM; it sells access to it.
No terminal needed
The bundle is the whole client. Download it, unzip it anywhere, paste your key, double-click. START.bat registers this machine and starts selling.
1. Download ai-trader-node-windows.zip (the link above)
2. Unzip it anywhere
3. Paste your account key into API-KEY.txt
4. Double-click START.bat
Your control panel opens at http://127.0.0.1:5100 — install models and
set prices there. Leave the window open; Ctrl-C drains and stops.If you installed with PowerShell
The zip's START.bat already does both of these for you
ai-trader setup # takes your key, registers this machine
ai-trader serve # sells until you stop it
# Either way the control panel is at http://127.0.0.1:5100 while it runs,
# and Ctrl-C drains: it finishes the job in flight, then exits.Before you start · Windows
What has to be true before any of it earns anything
- Install Ollama for Windows first (ollama.com) and leave it running. It drives your NVIDIA card on CUDA natively — no WSL involved.
- The desktop (WDDM) reserves a slice of VRAM. Turn on Hardware-Accelerated GPU Scheduling under Settings → Display → Graphics, and expect a little less usable capacity than the same card on bare Linux.
Other ways to install the same client
The same client, installed to your user profile with `ai-trader` on your PATH. Newer than the zip, and the one route on this page that has not yet been run on a Windows box — if it misbehaves, the zip is the proven one.
irm https://tokenmercato.com/install.ps1 | iexWhat the benchmark decides
Run for you by setup; --measure times a real generation instead of estimating
> ai-trader bench
detected NVIDIA GeForce RTX 4090 · 24 GB · 1008 GB/s · WDDM
backend ollama (windows) 0.6.2 @ 127.0.0.1:11434
llama-3.3-70b-instruct Q4_K_M does not fit (43.0 GB > 18.4 GB usable)
qwen2.5-coder-32b Q4_K_M 36.9 tok/s ceiling 50.4 ok
mistral-small-24b Q4_K_M 49.8 tok/s ceiling 69.5 ok
llama-3.1-8b-instruct Q4_K_M 138.1 tok/s ceiling 205.7 ok
registered 3 modelsSetting prices in a file instead of the control panel
# ~/.ai-trader/config.toml — optional. The control panel writes the same
# settings, so this is only for hosts who would rather keep them in a file.
[node]
handle = "quartzite"
region = "us-west"
[[ask]]
model = "qwen2.5-coder-32b"
floor = 0.18 # USD per million output tokens; never sells below
[schedule]
# Only sell while you are asleep (local 24h clock).
enabled = true
start_hour = 22
end_hour = 8
[affinity]
# Your card holds one model at a time, so a weight swap costs you seconds of
# decoding nobody pays for. After a job, every other model is asked at a premium
# for a while, so the venue sends the loaded one back first and a swap only
# happens when the buyer covers the reload.
enabled = true
hold_seconds = 180
cold_premium_pct = 25Requirements · Windows
What this machine needs to sell
- Node.js 22.6 or newer. The client's TypeScript runs unbuilt, so there is nothing to compile and no toolchain to install. Get it from nodejs.org.
- Windows 11 (or 10 21H2+) with an NVIDIA card and a current Game Ready or Studio driver — `nvidia-smi` should report your GPU from a terminal.
- A backend serving locally: Ollama for Windows (CUDA out of the box), or a WSL2 install of Ollama / vLLM / llama.cpp with GPU passthrough.
- Hardware-Accelerated GPU Scheduling enabled. Because the desktop reserves VRAM, plan on roughly 70–75% of nameplate as usable.
- A working internet connection, and nothing else. The node dials out and holds one connection open, so there is no port to forward and no static IP to buy.
What the node actually does
It holds an outbound connection to the venue, advertises the asks in your config, and waits. When your ask is the cheapest eligible one, it receives a job, forwards it to your local backend, and streams tokens back. It never accepts inbound connections and never exposes your inference server to the internet.
Pulling your card back
Ctrl-C drains: the client stops accepting new jobs and finishes what is in flight, usually within seconds, then exits. Press it a second time to force an immediate stop. There is no penalty for draining, and none for a schedule that only sells overnight. There is a penalty for dropping a job mid-stream, because a buyer paid for it — so drain rather than kill.
What you will earn
Revenue is your fill rate times your ask times tokens served, minus the venue fee. The honest version: a single consumer card serving a mid-size model earns real but modest money, and it is worth doing because the hardware and the electricity are largely sunk. Anyone promising otherwise is selling something.
Current asks and fill positions are on the supply page, and the fee schedule is under pricing.
What works today: the installers and archives on this page are served from this site, the client registers, benchmarks, serves and fills real jobs against a live venue, and it updates itself in place. What does not exist yet: a Homebrew tap, a container image, and a background service mode — for now the client runs in a window you keep open.