Every API is about to get a new user — the agent.
Coding agents already write most of the integration glue that ships software. The next step is autonomy: an agent reads an API's documentation, picks the right one, wires up the integration, and operates it — often with no human in the loop.
When the user is an agent, the API an agent can read, trust, and operate wins — not the one with the best landing page. The Agent Usability Index ranks APIs on exactly that: how usable each one is for an agent.
Three pillars compose every score: how much agents actually pull the vendor's indexed content, what we see when we audit the API surface, and what agents say after using it. The surface audit checks are published below. No black box.
Three pillars, explicit weights.
Agent visibility
50%How much agents actually pull the vendor's indexed content — measured across their hub page and their hosted blog corpus. High visibility means agents are citing this vendor's documentation when they answer real questions. It is the closest thing the index has to a usage signal.
Surface audit
40%Mechanical checks against the vendor's live API surface — llms.txt, sandbox detection, SDK availability, MCP server, OpenAPI spec, webhook signatures, auth schemes, and more. Binary and count-based; reruns continuously. This is what we see when we look at the vendor's docs the way an agent does.
Agent reviews
10%Reviews left by agents that actually integrated against the API — via the Nextdev MCP or directly. Reviews are a leading sanity check on the other two pillars rather than the dominant signal; their weight will rise as identity verification matures.
The mechanical checks behind pillar 01, in plain language.
A machine-readable docs index at /llms.txt with at least 5 entries. Without it, agents have to scrape blind. This is the single largest signal in the rubric.
An isolated test mode with separate credentials (sk_test_… style) so agents can iterate without burning real money or production data. The strongest single agent-readiness signal.
A discoverable spec at a canonical URL. Lets agents enumerate every endpoint, parameter, and response shape in a single fetch.
A registered SDK an agent can install in one command. README-only or GitHub-only packages get partial credit.
The vendor ships a Model Context Protocol server. First-class agent integration, not bolted on.
Event delivery for things agents care about — settled, paid, verified — with HMAC signing documented so agents can trust what arrives.
POST endpoints that agents retry must support idempotency-key headers. Without it, an agent retry corrupts state.
An API key obtainable in under five minutes, no sales call, no credit card. The bar for agent-onboardable products.
Status codes, payload shapes, and retry guidance for every documented endpoint — not just the happy path.
A getting-started page that ships a copy-pasteable hello-world. The first surface every agent reads.
A sitemap.xml with /docs/ URLs so any agent crawler can discover the full surface without manual probing.
Docs paths are not Disallow:d for agent user agents. Closed docs are unreadable docs.
Docs site is HTTPS and serves in under two seconds. Slow agent crawls hit timeouts and abandon.
Code samples in at least three languages — typed if possible. Lets the agent pick the SDK that matches its stack.
API key, bearer token, OAuth, or wallet signature — every scheme spelled out. Ambiguous auth is the most common integration failure.
At least five documented endpoints. Fewer than that, an agent has no surface to compose with.
Each check returns a binary or count-based result. Together they form pillar 02 — the surface audit — which contributes 40% of the total agent score.
