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

# Use with agents

> Lightdrift is built to be a tool in an agent's hands.

Lightdrift is agent-native by design: requests are plain JSON, every answer is fully machine-readable, and the rights question — the thing that usually needs a human — is answered inline.

## MCP server

<Note>
  The hosted Lightdrift MCP server ships with the beta. This page will carry the connection instructions for Claude, Cursor, and other MCP clients when it goes live.
</Note>

The MCP server exposes the same capabilities as the REST API as tools:

* `search` — natural-language text search with filters and negative constraints
* `find_similar` — expand from an asset you already have
* `get_asset` — full metadata and rights for one asset

## Patterns that work well

* **Grounding generation** — retrieve a real photo of an obscure subject first, then hand it to your image model as a reference. The generator stops guessing.
* **Slide and page builders** — ask for the subject *and* the layout: "room for a headline on the left" becomes a negative-space filter.
* **Answering with real images** — for anything past the canonical middle (a larva, a rare species, an obscure place), retrieval beats recall.
* **Rights-aware pipelines** — filter on `rights.commercial` and carry `rights.attribution` through to the final artifact automatically.

## Direct REST use

Agents that don't speak MCP can call the [REST API](/api-reference/introduction) directly — it's four endpoints, all JSON in and out. Image and multimodal query inputs are on the beta roadmap.