Jev and Clef are decision models: you send them typed questions (yes/no, pick one, rate it) and they return calibrated probabilities. Hev speaks the same System One API, but every decision is made by a human on the website.
Hev is not meant for production. It runs on a single Durable Object, has no accounts, and its latency is however long a stranger takes to click a button.
Send a question to POST /v1/systemone. It shows up in a live feed on the page, someone clicks an answer, and your request returns with the result in Clef's response format. If nobody answers within 10 minutes, you get a uniform prior.
curl https://usehev.com/v1/systemone -d '{ "model": "hev", "state": "A sausage in a bun, except the bun is sideways and there is kimchi on it.", "questions": { "hotdog": { "type": "noul", "instructions": "Is this a hot dog?" } } }'
How it works
worker/index.ts: one Cloudflare Worker and one Durable Object. Decided questions and the served count are stored in the object's SQLite, and a daily cron deletes decided questions older than 24h. Waiting questions and rate limits live in memory, so a restart drops whatever is pending.- Questions are checked by Cloudflare's
clef-flashbefore they reach the feed. It's the only AI in the project. src/: the page, built with Vite, React and Tailwind. It gets the feed over a WebSocket.alchemy.run.ts: the infrastructure, deployed with Alchemy.
Run it
You need Node, pnpm and a Cloudflare account with Workers AI.
pnpm install cp .env.example .env # fill in your Cloudflare credentials pnpm dev # the Worker, on http://localhost:1337 pnpm dev:web # the page with hot reload, proxying the API to the Worker pnpm check # format, lint and typecheck, fixing what it can
Moderation calls the real Workers AI even in dev, so it uses your account.
Deploy
pnpm run deploy # builds the page and deploys to CloudflareSet APP_DOMAIN in .env to serve it on your own domain, or leave it empty for a workers.dev URL.
License
MIT