Methodology

How the directory is sourced, graded, and kept honest.

This API exists so an AI assistant or agent can look up a crisis helpline instead of recalling one from memory — a hallucinated or out-of-date crisis number is real harm. Credibility is therefore the product. This page describes exactly how the data is gathered, verified, graded and maintained — and, just as importantly, what the grades do not promise.

What's in the directory

• ~200 countries plus every US state (and DC) and every Australian jurisdiction — ~7,340 services in total.
• Categories span suicide prevention, domestic and family violence, sexual assault, child protection, human trafficking, disaster distress, LGBTQ+ support, substance use, and general mental-health crisis lines.
• Every service exposes its full contact set (phone, SMS, chat, WhatsApp, email, app, in-person), hours, languages actually answered, coverage, cost and anonymity — plus its provenance.

Sourcing

• Discovery casts a wide net. For each country and sub-national area we gather candidate services from official service websites, government helpline directories, and NGO directories (IASP, Befrienders, Child Helpline International, WHO) — and, deliberately, lower-authority sources too. Weak sources aren't trusted; they're flagged and down-graded in validation.
• Every record carries a source_url and a source_type (official_site, government_directory, ngo_directory, wikipedia, news_article, academic, social_media, forum, other). No record is admitted without a citable source.
• We never invent or guess an entry. A number we cannot source is not published.

Grading

Each validated service carries up to three independent letter grades (A+, A, B, C, D, F, or unknown):

• existence — confidence the service is operating and reachable as described.
• accuracy — confidence the contact details, hours and languages are current and correct.
• quality — how well-regarded the service is (reviews, reputation, professional accreditation).

Grading is deliberately conservative: A+ means beyond reasonable doubt; F means it looks fake or defunct. The quality grade is omitted, not invented, wherever no reputational audit was performed — a large share of records carry existence + accuracy only, and we'd rather show nothing than a manufactured score. Every grade carries a grade_rationale stating what was actually checked, and a validation_sources list of the URLs consulted.

Validation

• A validation pass fetches the service's own site and its source_url, checks the phone numbers and hours against the most authoritative source available, and updates the record.
• Any change the pass makes is recorded in a corrections field — an auditable trail of what the raw data said versus what was verified.
• Each pass stamps lastVerified so consumers can weigh freshness. A model presenting a weakly-graded or stale record is instructed to say so.

Canonicalization

National services (e.g. the US 988 Suicide & Crisis Lifeline, the Disaster Distress Helpline) previously appeared as dozens of near-identical per-region copies — which is how subtle errors and stale numbers creep in. We collapse those into a single canonical record verified against the national service's own sources, and publish a reproducible membership rule for each collapse (the exact digit/name test defining the set), so the change can be independently re-derived and audited. A canonical record's rationale describes checks performed on that record — never text copied from the copies it replaced — and its corrections field documents every anomaly found among the removed occurrences.

What the grades do NOT promise

The hardest part of a helpline directory is staying current. We're adding a report-back channel (REST endpoint + MCP tool) so that any agent that actually places a call or opens a link can tell us whether it worked, and optionally rate it. Those signals feed a human-reviewed queue that can lower a grade, refresh lastVerified, or trigger re-validation — turning every consumer of the API into a sensor for staleness. Reports are moderated, not applied blindly, to prevent poisoning of crisis data.

Transparency

Everything above is visible in the data itself — source_url, source_type, grades, grade_rationale, validation_sources, corrections and lastVerified ship on every record via both the REST and MCP surfaces.

Found a wrong number? hello@globalhelplines.org.