# Use with coding agents > Connect these docs to Claude Code, Codex, Cursor, or any LLM so it writes your integration against the real contract. Source: https://docs.amomic.in/agents-setup --- If you are building this integration with a coding agent, give it the documentation first. An agent that has read the real contract writes working code on the first attempt; an agent guessing from a model's memory of "some WhatsApp API" invents endpoints that have never existed. Everything on this page is public and needs no key. ## The one-line version Pick whichever fits your tool: ::: codegroup ```bash Claude Code claude mcp add --transport http amomic https://docs.amomic.in/mcp ``` ```bash Any agent, any tool # Point it at the whole corpus — one fetch, everything. https://docs.amomic.in/llms-full.txt ``` ```bash Single page # Append .md to any docs URL for clean markdown. curl https://docs.amomic.in/api/messages.md ``` ::: ::: note The MCP endpoint `https://docs.amomic.in/mcp` is the documentation MCP server. It exposes these pages as searchable resources so an agent can query them mid-task instead of ingesting the whole corpus up front. If your tool does not speak MCP, `llms-full.txt` gets you to the same place with a single fetch. ::: ## What is available ```params [ {"name": "/llms.txt", "type": "curated index", "required": false, "description": "Every page with a one-line summary, organised by section. Small enough for any context window. Start here."}, {"name": "/llms-full.txt", "type": "complete corpus", "required": false, "description": "The entire documentation set in one plain-text file, in reading order. One fetch and the agent knows everything."}, {"name": ".md", "type": "raw markdown", "required": false, "description": "Any page, plus .md. /api/calls → /api/calls.md. No navigation, no HTML, no chrome."}, {"name": "/openapi.json", "type": "OpenAPI 3.1", "required": false, "description": "The machine-readable API contract, generated from the running service. Feed this to a client generator or an agent that speaks OpenAPI."}, {"name": "/mcp", "type": "MCP server", "required": false, "description": "Live documentation search over Model Context Protocol, for agents that support it."} ] ``` ## Set it up properly ::: steps ::: step Give the agent the entry point Whatever tool you use, the first thing it needs is a URL it can fetch. In a prompt, this is enough: ```text Read https://docs.amomic.in/llms-full.txt — that is the complete Amomic-AI documentation. Use it as the source of truth for endpoints, request shapes, and error codes. Do not guess at any API surface it does not describe. ``` That last sentence matters more than it looks. Without it, an agent that cannot find something tends to invent a plausible endpoint rather than tell you it does not exist. ::: ::: step Pin it in your repository Add the pointer to your repo so every future session starts informed, rather than you re-pasting the URL each time. ```markdown title="AGENTS.md" # Amomic-AI integration This service sends WhatsApp messages and places AI voice calls through Amomic-AI. **Documentation:** https://docs.amomic.in - Full corpus for ingestion: https://docs.amomic.in/llms-full.txt - Any page as markdown: append `.md` to its URL - OpenAPI contract: https://docs.amomic.in/openapi.json **Rules for this integration** - The public API is `https://services.amomic.in/api/v1`. Server-side only — it sends no CORS headers, so it must never be called from browser code. - Auth is `Authorization: Bearer amk_live_…`, from a secret store. Never a literal in source, never in the front end. - Every POST that spends money (`/messages`, `/calls`, `/tickets`) must send an `Idempotency-Key` derived from our own record id, so a retry cannot double-send. - Errors are `{"error": {code, message, details, request_id}}`. Branch on `error.code`, then `error.details.reason`. Log `request_id`. - WhatsApp free-form text only works inside the 24-hour window. Outside it, send an approved template or the API returns 409. - Phone numbers must match `^\+?\d+$` — no spaces, dashes, or brackets. ``` The same file works as `CLAUDE.md`, `.cursorrules`, or whatever your tool reads. The content is what matters, not the filename. ::: ::: step Give it the OpenAPI contract for exact shapes Prose is good at *why*; a schema is better at *exactly which fields*. For request and response shapes, point the agent at the contract: ```bash curl https://docs.amomic.in/openapi.json ``` It is generated from the running service, so it cannot drift from what the API actually accepts. ::: ::: step Check its work against the real thing The fastest way to catch a hallucinated endpoint is to call the API. `GET /ping` is free, has no side effects, and confirms auth, environment, and scopes in one response: ```bash curl https://services.amomic.in/api/v1/ping \ -H "Authorization: Bearer $AMOMIC_API_KEY" ``` If an agent has written code against an endpoint that does not exist, this is where you find out — in seconds, not in production. ::: ::: ## Per-page hand-off Every page here has a **Copy for LLM** button in its top-right corner. - **Click it** to copy that page as clean markdown, ready to paste. - **Open the dropdown** to send the page straight to Claude or ChatGPT with the canonical URL already in the prompt, or to grab the raw `.md` file. For a focused task — "implement webhook verification" — one page is better context than the whole corpus. Sending less, and sending exactly the right part, produces better code. ## Getting good results ::: tip Name the constraint, not just the goal "Send a WhatsApp order update" gets you a request. "Send a WhatsApp order update that survives a retry and handles the closed-window case" gets you the `Idempotency-Key` and the template fallback — both of which are in the docs, and neither of which the agent will reach for unless you say the words. ::: A few things worth telling any agent up front, because they are the mistakes that survive review: - **This API is server-side.** There are no CORS headers on `/api/v1`. If the generated code calls it from a React component, it will fail in the browser and the fix is architectural, not a header. - **201 on a call means ringing, not connected.** Outcomes arrive later, by polling or by the `call.ended` webhook. Code that treats the 201 as success will report every unanswered call as a win. - **Verify webhooks against the raw request body.** Parsing the JSON and re-serialising it changes the bytes and the signature will not match. This is the single most common webhook bug, in every language. - **Nothing is versioned behind a header.** There is one namespace, `/api/v1`. If an agent writes `Accept-Version` or `/api/v2`, it is inventing. ## Where the risk actually is Documentation makes an agent accurate about *our* API. It does not make it careful about *your* system. Two things still need human eyes: - **Where the key lives.** An agent will happily commit a working key into a config file because the code runs afterwards. Keys belong in a secret manager, and a leaked one should be rolled immediately from the console. - **Who gets contacted.** Code that loops over a customer table and calls `POST /messages` is trivial to generate and expensive to get wrong. Rate limits will slow it down; they will not tell you that you just messaged the wrong segment. Review those two by hand, every time. The rest of the integration is exactly the kind of work an agent should be doing for you.