> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sdk.anghami.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Features

> What AI surfaces the SDK exposes today, and what's coming.

The SDK is **agent-native by design** but does not yet ship server-side AI features (recommendations, summarization, agentic workflows). What it ships today is everything an external agent needs to consume the API correctly.

## What you get today

### Self-describing API

* **Bundled OpenAPI 3.1** ([`/api/anghami-sdk.openapi.yaml`](/api/anghami-sdk.openapi.yaml)) — every service, every operation, every type, in one document. Every field has a description (the proto's COMMENTS lint rule enforces this), so codegen and LLM grounding both work without manual annotation.
* **`llms.txt`** at [`/.well-known/llms.txt`](/.well-known/llms.txt) — concise briefing for LLM context windows. See [`llms.txt`](/llms-txt).
* **OAuth metadata** at `https://sdk.anghami.com/.well-known/oauth-authorization-server` (RFC 8414) — agents configure auth from one fetch.
* **Mintlify Ask AI** (this site) — documentation chat grounded in the MDX prose plus the OpenAPI bundle.

See [Agent Discovery](/agent-discovery) for the discovery surface.

### Patterns suited to agents

* **Cursor pagination only.** Agents iterating a catalog never deal with offset math or shifting result sets. See [Pagination](/usage-pagination).
* **Typed IDs.** `SongID` vs `AlbumID` vs `EpisodeID` — agents that mix entity kinds get type errors at codegen time, not runtime surprises.
* **Batch endpoints.** `BatchGetSongs`, `BatchGetEpisodes`, etc. — let an agent fetch the data behind a list with one tool call instead of N.
* **Idempotent reads.** All catalog/search/browse RPCs are safe to retry. Agents don't need to track "did this call succeed."
* **Structured errors.** `ErrorCode` enum + `field_violations` with paths — agents can self-correct on `INVALID_REQUEST` without parsing free-form messages.

## What's coming

* **Native MCP server.** A first-class MCP endpoint that exposes the same operations as tools, with `/.well-known/mcp/server-card.json` discovery. Tracked alongside the server implementation, not blocking SDK consumers today.
* **`api-catalog` linkset.** RFC 9727 discovery linking OpenAPI, MCP card, OAuth metadata, and `llms.txt` from one fetch.
* **Agent skills.** Pre-packaged recipes for common workflows ("get a soundtrack for a movie", "list a user's most-played artists last 30 days") that agents can ingest without reading the OpenAPI.

## What we're explicitly **not** building (yet)

* Server-side recommendation or summarization endpoints. The SDK is the data layer; if you want to summarize a user's library or generate a playlist with an LLM, you do that on your side using your own model.
* Hosted prompts or hosted tools. Bring your own model and prompt; we'll keep the data clean and the discovery surface honest.

## TL;DR

If you're integrating an agent today, the SDK already has what you need: a bundled OpenAPI, an `llms.txt`, and an OAuth flow. If you're waiting for native MCP — it's coming with the server, and when it ships, you won't need to redo your integration: the same operations and the same data, just exposed as MCP tools.