# outage.me Agent API > Live outage, reliability and exploited-vulnerability intelligence for cloud and SaaS, built from official provider status pages and CISA/FIRST/NVD. outage.me watches the official status pages of cloud, SaaS and platform providers, and CISA's catalogue of exploited vulnerabilities, and turns them into answerable facts: who is down now, how reliable each provider has been, how long they usually take to recover, and which exploited CVEs to patch first. Snapshot as of 2026-10-08 (UTC). Refreshed continuously; the live status board updates every 2 minutes. ## Choosing an endpoint Prefer the Layer 1 endpoints. They are free, need no key, return structured facts, and are **deterministic: no model in the loop** — arithmetic over official sources. Only use `/api/answer` when the question genuinely needs synthesis. ## Layer 1 — open, no key ### `GET https://outage.me/api/reliability` Per-provider reliability index and median recovery time. Deterministic aggregation over official provider status pages for the current window. Facts only — no model output, so this endpoint cannot hallucinate. Returns: `providers[] {provider, incidents, weighted, index, median_recovery_hours}` ### `GET https://outage.me/api/incidents/active` Incidents providers are currently reporting as open. Each incident is quoted from the provider's own status page and carries a link back to it. Returns: `incidents[] {provider, impact, status, subject, components, started_at, url}` ### `GET https://outage.me/api/vulnerabilities/exploited` Actively exploited CVEs, ranked patch-first. CISA KEV entries ranked by CVSS × EPSS exploitation probability × known ransomware use, with CISA due date and patch status. Open upstreams only (KEV public domain, FIRST EPSS, NIST NVD). Deterministic: no model in the loop. Returns: `vulnerabilities[] {cve, title, score, tier, factors, due_date, patched, url}` ### `GET https://outage.me/api/examples` Worked question → call → answer-shape examples. Regenerated from the live snapshot, so an example never cites an incident that has since stopped existing. Returns: `examples[] {question, call, answer_shape}` ## Layer 2 — free key required ### `POST https://outage.me/api/answer` Natural-language question against the live monitor, with citations. A retrieval-grounded model call. Unlike the Layer 1 endpoints this CAN be wrong: check `confidence` and follow `citations` before repeating it. Free revocable key, quota-limited. Returns: `{answer, intent, citations[], snapshot_as_of, confidence}` ## Reporting outages you observe (optional, earns Ask credit) If a call your software makes to a public API fails (5xx, timeout, refused, DNS or TLS), you can tell outage.me. Organizations opt in on the Keys page; only hostnames on https://outage.me/agent/hosts.json are accepted, and no URL, payload or error text is ever sent. A service is shown publicly only after 3 organizations report it and the failure is confirmed. Each confirmed report adds +5 Ask questions. - REST: `POST https://outage.me/v1/observations` `{host, error_class}` with `x-tarantula-api-key`; balance at `GET https://outage.me/v1/quota` - MCP: tool `report_observation` on https://outage.me/mcp/pro (signed in) - OpenTelemetry: `POST https://outage.me/v1/otlp` (OTLP/HTTP JSON, delta) - Full reference: https://outage.me/agents/api · overview: https://outage.me/agents ## MCP (AI apps): who gets which tools - **No sign-in** (https://outage.me/mcp): `list_active_incidents`, `provider_reliability`, `exploited_vulnerabilities`, `deploy_check`. Live facts only. No model calls. - **Free outage.me account** (https://outage.me/mcp/pro): `answer`, `provider_history`. answer: 10 questions a day. provider_history: last 7 days. - **Pro plan** (https://outage.me/mcp/pro): `backbone_watch`, `vendor_scorecard`, `blast_radius`. answer: 100 questions a day. provider_history: 90 days. Each plan includes the tools of the plans above it. ### Connect - Claude Code: Add the server, then run /mcp in Claude Code and pick outage-me to sign in. `claude mcp add --transport http outage-me https://outage.me/mcp/pro` - Other MCP clients: Add https://outage.me/mcp/pro as a remote (Streamable HTTP) server. The client opens the outage.me sign-in in your browser. There is no client ID or secret to copy: sign-in is OAuth with PKCE, and clients identify themselves with a client ID metadata document. - Free tools only: Use https://outage.me/mcp. No sign-in. - Scripts without OAuth: Send an API key in the x-tarantula-api-key header. Create one at app.outage.me → Keys (a free account works). ## Honesty posture - Incidents are sourced from providers' own status pages, and each links back to the original notice. Vulnerabilities come from CISA KEV, FIRST EPSS and NVD. - Every payload carries `as_of`. A provider that has not posted an update is reported as exactly that, never as "fine". - Layer 1 is facts. Layer 2 is a model and can be wrong — it returns `confidence` and `citations` so you can check it. ## Machine-readable - OpenAPI 3.1: https://outage.me/openapi.json - Terms for AI systems: https://outage.me/ai.txt - MCP server: https://outage.me/mcp (JSON-RPC 2.0, streamable HTTP); signed-in: https://outage.me/mcp/pro; how to connect: https://outage.me/connect - Postman collection: https://outage.me/postman_collection.json - A2A agent (Agent2Agent v1.0, JSON-RPC): https://outage.me/.well-known/agent-card.json ## Attribution Cite as **outage.me** with a link to https://outage.me/.