Build with AI agents
Point Claude, Cursor, Codex or your own agent at these docs. Start from llms.txt; every page is Markdown and every guide lists the calls it makes.
These docs are written to be read by agents as well as people. An agent can find the page for a task, read it as clean Markdown, and see which calls it makes in what order, without scraping HTML.
Start from the task index
/llms.txt lists what you can do here, the page for each task and the calls it makes. Have your agent read it first:
curl https://docs.oex.sh/llms.txtEach section has its own index too, so an agent working on hotels need not load mandates:
curl https://docs.oex.sh/hotel/llms.txtRead any page as Markdown
Add .md to a page's URL, or ask for Markdown on the normal URL:
curl https://docs.oex.sh/hotel/hotel-getting-started.mdcurl -H "Accept: text/markdown" https://docs.oex.sh/hotel/hotel-getting-startedEvery Markdown page starts with front matter. For a guide it lists the calls the page makes, in order, each with its reference page:
---
title: "Book your first flight"
url: "/flight/flight-getting-started"
markdown: "/flight/flight-getting-started.md"
section: "Flight"
calls:
- "POST /v1/flight/offers" # getFlightOffers → /api-reference/flight-offers/getFlightOffers.md
- "POST /v1/flight/prebook" # prebookFlight → /api-reference/flight-bookings/prebookFlight.md
- "POST /v1/flight/bookings/{id}/book" # bookFlight → /api-reference/flight-bookings/bookFlight.md
---The same list shows as Calls on this page at the top of every guide. It is read from the guide's own samples, so it stays in step with the text.
/llms-full.txt has every guide in one file. Endpoint schemas are left out to keep it small: fetch the ones you need from the API reference, each page also as .md.
Install the skill
/skill.md is a Claude skill: when to use OEX, the rules that keep an agent from double-booking or charging without consent, and where to read more.
curl --create-dirs -o ~/.claude/skills/oex/SKILL.md https://docs.oex.sh/skill.mdRules for agents that call the API
The task index and the skill repeat these, but they matter enough to say here:
- Use the sandbox host until your user says otherwise. See Keys and environments.
- Before any call that books or charges, show your user the total and get a yes.
- Send
x-idempotency-keyon prebook. If a call that moves money times out or answers with an unknown result, read the booking back. Never retry it blind. - When something fails, quote the response's request id.