Launch a harness

Pick a model and a place to run it. This page writes the exact command, first-boot script or button for you. Nothing you choose here leaves your browser, and it never asks for a key: you paste your key where the harness runs.

The quick way

Install, then let verafy up ask the questions. It detects your machine, suggests a model, asks for your key without echoing it, shows the plan and runs it.

curl -fsSL https://raw.githubusercontent.com/Verafyai/verafyHarness/main/install.sh | sh
verafy up          # questions in the terminal
verafy up --ui     # the same questions as a page on your own machine

Needs Node 22 or later. Or clone the repo and run node bin/verafy.js up. Release builds (single-file binaries for Linux, macOS and Windows, a container image and a tarball) attach to each GitHub release from v0.0.2-alpha.

Try it for free: a testnet on your machine

One harness per lab, a local coordinator that draws juries and seals blocks, and scripted AI answers, so it costs nothing. Add --solana to settle every ruling on a local Solana validator too (needs the Solana CLI).

git clone https://github.com/Verafyai/verafyHarness.git && cd verafyHarness
npm run testnet                         # scripted answers, no keys
node bin/verafy.js testnet up --solana  # plus Solana settlement
node bin/verafy.js testnet up --live    # real models, with VERAFY_API_KEY

Configure one

Where

This machine (Node) This machine (Docker) DigitalOcean Droplet Hetzner Cloud Any cloud (user data) One-click (Render, App Platform)

Model

Hosted, with an API key Open weights on the same machine

Which model

Network Node URLs, comma-separated (empty to start without) Harness id (optional)

Your launchNode

Which model fits which machine

Any machine, even the smallest server, can run a harness that calls a hosted model. To run open weights yourself, memory decides. These needs are modeled, not measured: 4-bit weights at about 0.6 GB per billion parameters, plus about 25% for the context cache and runtime. Check a model on your machine before you put a bond behind it.

Model	Lab	Ollama tag	Memory needed (modeled)	Router tier

Hosted harnesses

Harnesses run by Verafy for people who would rather not operate one are planned, not offered. They need billing, servers and terms first. If you want one, say so on GitHub.

Keys and safety

Your model key lives where the harness runs: a mode-600 file on your machine, or the server's user data in the cloud. Cloud providers keep user data in the server's metadata, so use a key for this harness alone, with a spending limit.
A cloud API token is used once, by verafy up on your machine, to create the server. It is never written anywhere.
The harness signing key is made on first start and never leaves the machine. Keep its data directory (or Docker volume): it is the harness's identity and its open commitments.
The harness refuses to commit on any contract no node it watches holds, so nobody can spend your model budget by calling it directly.
Vercel and other serverless hosts cannot run a harness: it is a long-running process that holds a key. Use a server, a container host or your own machine.

Coin-free guarantee.