Give your AI the API contract
The OpenAPI 3.1 specification is the single source of truth — every endpoint, schema, auth requirement, and status code:
In Claude Code, Cursor, or any agent with web access, this is usually enough:
Prefer the SDKs over raw HTTP
Ask the AI to use the official SDKs (hostpay
on PyPI, @hostpay/sdk on npm)
rather than hand-rolled requests. Both are typed, so AI-written code that
misuses the API fails your typecheck instead of failing in production —
the tightest feedback loop an agent can get.
Starter prompt
Paste this at the top of your session, filling in your keys:Verify what the AI built
Don’t trust — run. Test Mode makes agent-written code cheap to prove:- Run it against Test Mode keys; the magic numbers make outcomes deterministic (see the Testing guide).
- The SDK repo ships a runnable end-to-end check
(
examples/smoke_sdk.py) you can point at your integration environment. - Typecheck:
mypy/Pyright for Python,tscfor TypeScript — the SDKs’ types catch hallucinated fields.
Grounding the docs themselves
Every page on this site is written in plain MDX; if your assistant supports web fetching, point it at the specific guide you’re implementing (e.g./guides/deposits, /webhooks/security) alongside the OpenAPI spec.