A human can paper over a bad API. Missing an example? Guess from the field name. No idempotency guarantee? Retry and eyeball the result. Auth documented in prose? Read it and move on. An agent cannot do any of that. It needs the conventions written down in a form it can parse. That gap is why APIs.io rates every provider on a standalone Agent Readiness score, kept separate from the composite quality score — not blended into it.
The separation is the point. A provider can be excellent for human developers and still be unsafe to hand to an agent. Folding the two together would hide exactly the signal we want to surface, so Agent Readiness stands on its own axis.
What it measures
Agent Readiness is a 0–100 score, normalized from 12 dimensions worth 104 raw points. The weighting says plainly what agents most depend on:
- Machine-readable contract (18) — the largest single award; no spec, no autonomy.
- Agentic-access contract / x-agentic-access (15) — an explicit statement of what an agent may do.
- MCP server (12) — a native tool surface an agent can call directly.
- Machine-readable auth (10) — credentials it can negotiate without a human.
- Idempotency (9) — safe retries.
- Stable error semantics (8) — failures it can reason about.
- Examples (7) and rate-limit signaling (7) — concrete payloads and backpressure it can respect.
- Typed event surface (6), agent skills (5), well-known catalog (4), and consent & bot-identity (3) round out the rubric.
The through-line: every point rewards replacing an implicit convention with an explicit, machine-readable signal.
The four bands, honestly
The distribution across the catalog is not flattering, and it is not meant to be:
- Agent-Native (60+) — roughly 1.4% of the catalog. The genuinely autonomy-ready.
- Agent-Ready (45+) — about 43%. Workable with human setup.
- Agent-Aware (15+) — about 16%. Some signals present, big gaps remain.
- Human-Only (0–14) — about 39%. Built for people reading docs, full stop.
Read together, that means only around one provider in seventy clears the top bar today, and nearly four in ten offer an agent almost nothing to work with. The score is a map of how far the catalog has to travel, not a victory lap.
Takeaway
Agent Readiness is a diagnostic, not a grade. It tells a provider precisely which implicit conventions to make explicit — publish the spec, declare agentic access, stand up an MCP server, document auth and errors in machine-readable form — and it tells a consumer which APIs an agent can actually be trusted to drive. As agents move from demos to production, this is the axis that will matter most. See how every provider scores at apis.io/rating/.