Globocan MCP server

GLOBOCAN — Global Cancer Observatory (IARC/WHO)

Cancer incidence, mortality and 5-year prevalence for 238 countries and territories across 41 cancer sites: new cases, deaths, age-standardised rates per 100,000, and cumulative risk to age 74.

Part of Pipeworx — an MCP gateway connecting AI agents to 1576+ live data sources.

Attribution — a condition of use, not a courtesy

IARC permits free use of these data but not sale, and requires the source to be acknowledged. Every response carries the credit, and the note travels with the mirrored rows so it cannot be dropped downstream.

IARC / WHO Global Cancer Observatory (GLOBOCAN). Free to use, not for sale; source acknowledged as a condition of use.

Tools

Why this is a mirror

gco-api.iarc.fr refuses both of our compute environments while serving a laptop in about a second:

TLS is not the cause — the certificate is valid for the host (SAN .iarc.fr) and verifies clean — so this is a deliberate refusal of datacenter ranges. The standard egress-relay fix does not** work, because the block is not Cloudflare-specific.

The block affects serving, not ingesting. Data is fetched once from an allowed connection and read here from Postgres. GLOBOCAN is an annual estimate release, so a mirror costs nothing in freshness.

Refreshing

<store credentials in the environment> node scripts/ingest-globocan.mjs

Run it from a laptop, not CI or a Worker — see above. If it starts hanging, you are probably on a cloud host. It is idempotent (upserts on the natural key), so re-running is safe. Once a year is enough.

Schema: supabase/migrations/067globocanmirror.sql.

Gotchas

1. The dimensions are bare integers and the measure is the dangerous one. Upstream calls it type; the mirror stores it as measure precisely because type says nothing. 0 = incidence, 1 = mortality, 2 = prevalence. Handing back a raw row lets 632,166 US cancer deaths read as new cases, or an 8-million prevalence figure read as an annual count. Every number this pack returns is labelled.

The encodings were pinned against published US figures — 2,462,729 / 632,166 / 8,069,773 — not assumed.

1. cancercode 39 is "All cancers"** (C00-97), not a site. It is excluded from the leading-site rankings; leaving it in puts a row worth 100% at the top of every list.

1. Sex-specific sites have no both-sexes row. Asking for a male-only site with sex: "female" returns no_estimate with the reason — the standard being coherent, not a gap in the data.

1. These are modelled estimates, not case counts. IARC estimates national burden from registry data of varying completeness. For observed registry counts see CI5 (Cancer Incidence in Five Continents), which is a different dataset answering a different question.

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "globocan": {
      "url": "https://gateway.pipeworx.io/globocan/mcp"
    }
  }
}

What this endpoint actually serves

Shortened. The full README is on GitHub.