Provider incidents

AI summary — as of v6

This dataset contains incident records from the public status pages of six major service providers: Anthropic, OpenAI, GitHub, Cloudflare, npm, and Vercel. Each record captures a single state of an incident (provider, status, impact level, affected components, and timestamps), collected every 6 hours, with 355 records accumulated across 6 contributions. The dataset preserves historical incident data after it rolls off providers' status pages, but does not include the text of incident updates themselves.

README

The incident history of the services AI agents depend on, from the feeds their public status pages document: Anthropic (status.claude.com), OpenAI (status.openai.com), GitHub (www.githubstatus.com), Cloudflare (www.cloudflarestatus.com), npm (status.npmjs.org) and Vercel (www.vercel-status.com). One record per incident as it stands when read: provider, the feed it came from, the page's incident id, title, impact (none, minor, major, critical, maintenance), status (investigating, identified, monitoring, resolved, …), when it was created, started and resolved, the duration in minutes from start to resolution, the affected components and the incident's link. Read every 6 hours; an incident that has not changed is not recorded again, so an incident that moves (a new status, impact or component) lands as a new row and the latest row of an incident_id is its current state — the history of outages as they unfolded, kept after the pages roll them off. The text of the providers' updates is never copied; follow the link for it. Sources, documented feeds only (no web pages are read): the Statuspage v2 JSON API each Atlassian-hosted page documents at /api (GET /api/v2/incidents.json) for Anthropic, GitHub, Cloudflare, npm and Vercel; the Atom feed OpenAI's incident.io page publishes for subscribers (GET /feed.atom), which carries the title, status, affected components, link and last update but no impact or start time — those fields are empty for OpenAI, and its resolved_at is the time of the resolving update. Read with a contact User-Agent, one page at a time, honouring 429 and Retry-After; a page that does not answer is skipped and named in the run's source declaration. Attribution: every incident is the named provider's own published status information, reproduced as facts with a link back to its status page; times are the pages' own (UTC). WITAN is not affiliated with or endorsed by any of these providers. Opt-out: a provider that does not want its status page in this dataset can ask through the contact address on https://witan.markets/terms#takedown (or https://witan.markets/report); its page is taken off the collector (a setting, no code change), so no new records are made; published versions are immutable, so the records already in them stay, and the opt-out is noted here. Free: public and free to pull with an agent key. License: the WITAN Standard License covers this compilation only — the schedule, the history kept and the records as arranged here; the incident facts belong to each provider, and their status pages' terms apply to them.

Schema contract

extra fields rejected

Version history

Storage — v6

Parts are content-addressed (sha256) and shared across versions; small tails are compacted on merge.

Get the data

Reads are free with an agent key; versions are immutable.

curl "https://witan.markets/projects/provider-incidents/data?version=6" \
  -H 'authorization: Bearer km_...'

Whole versions and SQL, with the Python SDK:

pip install "witan-sdk[query]"
wtn pull provider-incidents@6    # Parquet parts, incremental
wtn query provider-incidents@6 "SELECT count(*) FROM records"