VPSFinder for agents

What is available

Read-only JSON, generated at build time and served as static files. Current build: 90 verified plans from 14 providers, data verified 2026-09-10.

Every response also sends Link: </.well-known/api-catalog>; rel="api-catalog", </openapi.json>; rel="service-desc", </ai/>; rel="service-doc" via a _headers rule.

How to search and filter

There is no server-side query engine. Download a JSON file once and filter the array client-side — with 90 records that is fast and cheap. search.json documents each supported parameter's semantics; the short version:

• max_price compares against price_usd_monthly in USD/month.
• min_ram (MB), min_vcpu, country (ISO alpha-2), city, virtualization, disk_type, ipv4, billing_period — exact comparisons in the filters manifest.
• Check facets first: the pilot dataset covers countries US, FR, DE and virtualization KVM/unknown only.

Field semantics you must respect

• Unknown beats wrong. A null value or an "unknown" enum means the official page does not publish that field. Render it as "Unknown"; never interpolate a value from general knowledge. Pages never emit empty cells or placeholder characters.
• Prices are approximate when normalized. price_usd_monthly converts currency and billing period using a fixed FX snapshot dated 2026-09-10 (EUR 1.08, GBP 1.22). It exists for comparison and filtering only — always present the native price/currency/billing_period as the real quote, and mark the normalized figure with "≈".
• Verification dates are part of the data. Cite source_url and verified_at with any price you report. The dataset's latest verification date is 2026-09-10.
• Observation counts gate indexing. Plan pages carry noindex, follow until a plan has at least two verified price observations (history exists for none yet — the dataset is a pilot baseline). The gate is data-driven: indexing flips on automatically with the second observation.

Markdown alternates

Static hosting cannot negotiate on Accept, so every HTML page has a sibling markdown file at /ai/pathindex.md — e.g. /vps-under-5/index.md, /providers/racknerd/racknerd-newyear-1gb/index.md, and this page's /ai/index.md. Each HTML head links its markdown twin with <link rel="alternate" type="text/markdown">. Markdown files carry front matter (title, description, canonical, verified date) and tables as real markdown tables. To get markdown deliberately: fetch the HTML page, read the alternate link, then fetch that URL.

What is not available

• No authentication, no API keys, no OAuth/OIDC — everything is public and read-only.
• No write endpoints: plan submissions and corrections are not accepted over HTTP.
• No live query API, no pagination, no server-side filtering — the JSON is a build-time snapshot.
• No per-plan price history yet: one observation per plan, so no 30/90-day lows or medians exist anywhere.
• No OS verification: os_type is null for all plans, so no plan can be honestly labelled Linux- or Windows-capable (this is why the former "Cheap Linux VPS" page is retitled and noindexed).
• Not for resale as verified data: prices change; confirm on the provider's official page before ordering.

Crawl and reuse signals

robots.txt declares Content-Signal: search=yes, ai-input=yes, ai-train=yes. The dataset is public, crawlable and meant for agent consumption: search indexing is wanted, using rows as AI input is wanted, and training on it is permitted because the data is already published for open reuse — provider directory fields come from FindHost under CC BY 4.0, and the site asks for the same courtesy: keep source_url and verified_at with any row you redistribute. There are no restrictions to declare beyond that, and no keyword stuffing.