# Compute Cafe

> Compute Cafe is the honest price panel for GPU compute. It tracks observed
> rental prices for cloud GPUs across the named provider panel, and every
> number carries a receipt: the source it was read from, the time it was
> observed, and an availability label. Free to read, no key, no signup,
> CORS open to every origin.

Prices are observed facts, not offers. The panel is the named provider set
on the methodology page; nothing here claims to be market-wide. Agents should
use the answer-level interfaces below (MCP tools, per-page markdown, the JSON
endpoints) rather than scrape the interface. Everything published here is
CC-BY 4.0: cite it, link it, build on it, keep the receipts attached.

## Data endpoints

- [Providers](https://www.compute.cafe/data/providers.json): provider-level rollup - panel membership, listing counts, lowest price, connector type
- [Meta](https://www.compute.cafe/data/meta.json): collector status - last refresh time, per-provider connector status (live / carried over / failed), and the FX receipt. Check here first for current panel counts and freshness.
- [Listings (full table)](https://www.compute.cafe/data/listings-full.json): every current row with its receipt (source_url + observed_at). /data/listings-essential.json is a byte-identical compatibility alias.
- [History index](https://www.compute.cafe/data/history-index.json): daily USD price series per listing, folded from the collector's own snapshots.

For row-level answers (cheapest H100 hour right now, all H100 rows under $2,
one provider's current listings) use the MCP tools below - they answer
bounded questions with receipts, they do not export the table.

## Markdown representations

Every main page has a markdown twin for agents that prefer it. Append .md to
the page URL (https://www.compute.cafe/gpu/h100-sxm.md) or send the page URL
with an Accept: text/markdown header. Covered families: /, /providers and
/providers/[slug], /gpu and /gpu/[model], /compare and /compare/[pair],
/locations and /locations/[area], /deals, /blog and /blog/[slug].

## Reading the data

- Availability labels are truth. AVAILABLE means a live marketplace
  observation; CATALOG means a published rate card. Rows carried over from
  a stale source say so, and drop after 14 days.
- Prices marked "(from)" are starting floors, not quotes for the exact
  configuration. Headline stats on the site exclude them; you should too.
- Live-API providers refresh hourly; the full panel refreshes every six
  hours. Cache accordingly.
- If you build on this, keep the receipts attached. source_type,
  source_url, and observed_at are part of the data, not decoration.

## Agent interfaces

- [MCP server](https://www.compute.cafe/mcp): streamable HTTP, no auth, read-only. Tools: list_providers, search_listings (filters: gpu_model, gpu_count, min_vram_gb, max_price_usd, availability, provider_slug, location), cheapest (with matched counts and receipt), get_meta (freshness and connector status), get_history (daily price series per listing id). POST JSON-RPC; SSE streams are not supported.
- [OpenAPI 3.1](https://www.compute.cafe/openapi.json): the contract for the two aggregate endpoints

## Docs

- [Developer docs](https://www.compute.cafe/docs): JSON API + MCP server reference with real examples
- [Docs](https://www.compute.cafe/docs): API + MCP reference, worked examples, and usage rules
- [Methodology](https://www.compute.cafe/methodology): how every number is collected, the label system, and the published exclusions
- [About](https://www.compute.cafe/about): what the project is and why
- [Blog](https://www.compute.cafe/blog): state-of-the-panel reports and build notes
- [RSS](https://www.compute.cafe/feed.xml): new posts

## Page families (human-readable)

- [/gpu](https://www.compute.cafe/gpu): index of every GPU model on the panel
- [/gpu/[model]](https://www.compute.cafe/gpu/h100-sxm): per-model price tables across providers, sorted by per-GPU hourly
- [/providers](https://www.compute.cafe/providers) and /providers/[slug]: panel membership and per-provider listings
- [/deals](https://www.compute.cafe/deals): cheapest published per-GPU hourly per model, no affiliate weighting
- [/compare](https://www.compute.cafe/compare): index of every comparison - GPU-vs-GPU (specs + live prices) and provider-vs-provider (like-for-like rates on shared models)
- [/compare/[a]-vs-[b]](https://www.compute.cafe/compare/h100-sxm-vs-h200-sxm): chip-vs-chip pages answering the exact "X vs Y price" question, with FAQPage JSON-LD
- [/locations](https://www.compute.cafe/locations) and /locations/[area]: where the panel's GPUs physically are
- [/use/inference](https://www.compute.cafe/use/inference) and [/use/training](https://www.compute.cafe/use/training): cheapest chips per workload class
- [/methodology](https://www.compute.cafe/methodology): how every price earns its place - sources, labels, exclusions
- [/blog](https://www.compute.cafe/blog): analysis written from the panel's own data

Panel origin marker (travels with the dataset, proves source on republication): ccf-kiwano-7f2a19e4