Agents already won: more public APIs ship an MCP server than an OpenAPI spec.

We verified the access terms of 1,340 public APIs — auth, pricing, free tiers, rate limits, specs, MCP servers — reading each vendor's own docs and storing a source URL for every value. Here is what the machine-readable layer of the API economy actually looks like in mid-2026.

This is the finding that reframes everything else. The incumbent open directory, apis.guru, spent its life indexing OpenAPI specifications — then froze in April 2023. But the OpenAPI spec turns out to be the rarest machine-readable artifact in the entire corpus: barely 7.2% of public APIs expose one at a discoverable URL. Meanwhile MCP servers — the interface built for AI agents — already appear on 2.25× as many. The tooling ecosystem is still organized around the spec. The APIs have moved on.

Free tiers are the default

Among the APIs whose terms we could fully verify, 91% document a free tier and 61% publish their rate limits. Free access isn't a growth hack anymore — it's table stakes. But the distribution by category is where it gets interesting: the APIs least likely to document a usable free tier are the ostensibly "free" ones — government and open-data APIs, where terms are so under-documented that whether you can use them in production is often unstated.

The auth surface is simpler than the tutorials suggest

Of the APIs that document authentication, the API key still rules — and a fifth need no auth at all. OAuth 2.0, the thing every integration guide dwells on, is the minority case. For anyone building an agent that calls arbitrary APIs, that's good news: the auth surface is mostly a single static credential.

Pricing has consolidated

Freemium and genuinely-free account for most documented pricing. Pure pay-to-play is rare; even in the metered-inference era, usage-based billing is still the minority. The modern default is settled: a free bucket, then usage or seats.

Only a quarter are machine-discoverable at all

Across every live API domain we probed, just 25% serve an llms.txt and 7.2% an OpenAPI spec URL. For most public APIs, an agent's only route to the terms is reading the human docs page — which is exactly the gap this census exists to close.