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

# MCP Overview

> Connect Datagate to AI agents via Model Context Protocol

## What is MCP?

[Model Context Protocol (MCP)](https://modelcontextprotocol.io) is an open standard that lets AI agents discover and call external tools. Datagate provides MCP servers that expose dataset discovery and querying as tools — no integration code needed.

## When to use MCP vs SDK

| Use case                                                 | Use                                            |
| -------------------------------------------------------- | ---------------------------------------------- |
| Claude Desktop, Claude Code, or any MCP-compatible agent | **MCP server**                                 |
| Custom Python/TypeScript application                     | **SDK**                                        |
| Building your own agent framework                        | **SDK** (or MCP if your framework supports it) |

## How the packages relate

```
datagate          ← SDK (standalone HTTP client)
   ↑ depends on
datagate-mcp      ← MCP server (thin wrapper around the SDK)
```

The **SDK** (`datagate`) is the foundation — a Python/TypeScript client that calls the Datagate API directly. It returns typed objects and gives you full control.

The **MCP server** (`datagate-mcp`) is a wrapper that imports the SDK and exposes its methods as MCP tools. When you install `datagate-mcp`, the SDK is automatically installed as a dependency.

You can use either one independently:

* **SDK only** (`pip install datagate`) — for custom apps and agents where you want typed responses
* **MCP only** (`pip install datagate-mcp`) — for Claude Desktop, Claude Code, or MCP-compatible frameworks (SDK comes along automatically)

## Available Packages

| Package         | Language   | Install                     | Run                    |
| --------------- | ---------- | --------------------------- | ---------------------- |
| `datagate-mcp`  | Python     | `pip install datagate-mcp`  | `uvx datagate-mcp`     |
| `@datagate/mcp` | TypeScript | `npm install @datagate/mcp` | `npx -y @datagate/mcp` |

Both expose the same two tools and produce identical output.

## Tools Exposed

### list\_datasets

Calls `GET /v1/datasets/discover`. Returns all visible datasets with:

* Name, ID, seller, description
* Price per chunk
* Queryable/subscribed status
* Metadata schema (available filter fields)

### query

Calls `POST /v1/query`. Parameters:

* `dataset_ids` (required) — which datasets to search
* `text` (required) — natural language query
* `top_k` (optional, default 10) — max results
* `filters` (optional) — per-dataset metadata filters

## How an Agent Uses It

1. Agent starts → spawns MCP server → discovers `list_datasets` and `query` tools
2. User asks "What datasets are available?" → agent calls `list_datasets`
3. User asks "Find privacy policy precedents" → agent calls `query` with dataset IDs and text, using metadata filters from step 2
4. Agent reads results and summarizes for the user

The MCP server always queries with text — Datagate handles embedding and multi-model orchestration automatically.