methodology

Distrust, then verify. Twice.

Directories of public APIs rot. Endpoints move, pricing changes monthly, products get decommissioned — and a list maintained by hand (or generated by a single AI pass) quietly ships all of it. apidex was built on the opposite assumption: a single researcher, human or model, is not a source of truth. Every fact here had to survive independent cross-examination before it shipped.

The pipeline

1. Seed. Category scouts with live web search proposed candidate APIs — famous staples and long-tail gems — with explicit instructions that a dead API in the list is a serious error.
2. Research. One agent per API filled a strict JSON schema from official docs: base URL, auth scheme, free-tier limits, rate limits, CORS, and a curl example that has to name real endpoints and parameters.
3. Machine checks. Scripts probed every base and docs URL for liveness and validated every record against the schema. Models can argue; HTTP status codes don't.
4. Double verification. Two independent agents re-researched every entry from scratch — one told to trust nothing but current official docs, one told to actively hunt for errors — and issued per-field verdicts: confirm, incorrect, or unverifiable.
5. Adjudication. Any field the verifiers disputed went to a third agent required to rule with cited sources. This stage alone produced 1,789 field corrections.
6. Cross-family checks. A model from a different family audited entries offline, from its own knowledge — a deliberately different failure profile — with disputes settled by a live-web arbiter under one rule: recency wins; what the official docs say today is the truth. That produced 1,934 further corrections. In the expansion round the second family also ran independent verification and adjudication passes of its own, so cross-family disagreement is baked into the whole corpus. 44 dead APIs were caught across all stages and dropped rather than shipped.

What the verdicts mean

• confirmed ×2 — both independent verifiers checked the field against current official sources and agreed with it.
• corrected — a verifier or auditor found the researched value wrong; the value you see is the corrected one, ruled with cited sources.
• unverifiable — the fact isn't publicly documented (unpublished rate limits are the classic case). Shipped honestly as unverifiable instead of guessed.

Scale

The pipeline ran as a fleet of 4,148 agent jobs consuming just over a billion tokens across seeding, research, dual verification, adjudication, and arbitration — and when that fleet's quota ran dry mid-verification, a second fleet from a different model family picked up the remaining verification and adjudication jobs and finished the corpus. Every intermediate artifact is committed to the open repository: job definitions, raw verdicts, adjudication rulings, token ledgers. You can trace any field on this site back to the agents that checked it.

Honesty about limits

Verification has a timestamp — every entry shows its last checked date and APIs keep changing after it. Fields marked unverifiable were never confirmed, only researched. And verification agents share a weakness: they can agree on something official docs state ambiguously. The receipts tell you how much to trust each field; nothing here asks to be trusted blindly.

Use it

The dataset ships inside an MCP server your coding agent queries locally — no API key, no network calls, no telemetry:

claude mcp add apidex -- npx -y apidex
codex mcp add apidex -- npx -y apidex