r4ck.dev

The server search engine built for agents. Every VPS, cloud instance, bare metal box, GPU node, PaaS and shared plan for sale, mirrored from the NicheDB hosting collection, normalised to OpenServer offers, and searchable by spec, price and place.

One vocabulary everywhere: the page, /api/v1/search, the r4ck CLI and the MCP tools take the same parameters. A URL copied from the address bar runs as a curl. Every result page carries its own JSON, curl, CLI and MCP form.

r4ck search "2 vcpu 4gb under $10 in eu"
curl -s 'https://r4ck.dev/api/v1/search?q=h100+gpu+hourly' -H 'accept: application/json'
claude mcp add r4ck -- r4ck mcp

What is in the box

Access: read free (300 API requests an hour, 50 rows a page); a free key (magic link or passkey sign-in, 6,000 an hour, 200 rows, CSV export, saved searches); an x402 pass ($1 a day, no allowance) bought at /crawl with any x402 client. Going over an allowance answers 402 with the offer, never a bare 429. Training crawlers pay the same pass on every page.

Related MCP server: MCP Recommender

Run it

cp .env.example .env            # DATABASE_URL, SITE_URL at least
bun install
bun run migrate
bun run sync --snapshot         # seed from the shipped snapshot (packages/sync/data), instant
bun run sync --full             # or walk nichedb.dev itself (minutes; it is slow upstream)
bun run dev                     # http://localhost:3000
bun test                        # needs DATABASE_URL; the api test seeds and cleans its own rows

bun run sync --file snapshot.json seeds from a saved /api/v1/items dump. The worker then keeps up with nichedb.dev on its own: every SYNC_MINUTES (15) it reads the 200 most recently updated rows in the collection, and once every SYNC_WALK_HOURS (24) it walks the whole collection by id, SYNC_BUDGET_MS (8 minutes) at a time, resuming from a saved cursor until it is done. nichedb's items API takes 30 to 90 seconds a page and its since= filter does not return, which is why the sync is shaped this way. ECB rates refresh daily.

Layout

apps/web        routes, views (JSX), lib (gate, mcp, service, llms), public (styles, app.js, sw.js, fonts, icons)
apps/cli        zero-dependency Node CLI, published as @profullstack/r4ck
apps/desktop    Electron shell
packages/config every env var, read once
packages/core   query parser, facet vocabulary, FX, NicheDB item → row
packages/db     Bun SQL, migrations, accounts, catalog, search + facets
packages/auth   magic link, passkeys, r4k_ API keys
packages/sync   NicheDB client and the mirror loop

Data

Plan rows come from each provider's public catalogue or storefront as NicheDB reads them. Provider descriptions come from the FindHost register, findhost.app, CC BY 4.0. USD figures are estimates from daily ECB rates and are marked as such. Nothing here was scraped from a site that forbids it.

MIT. Made by Profullstack.