# Core Concepts

> The foundational ideas behind Alpha Ladder's MCP-first architecture — what MCP is, the AgentX product, tool safety, and the first-party vs. third-party model.

## MCP, in one paragraph

[MCP (Model Context Protocol)](https://modelcontextprotocol.io) is a standard that lets AI clients call external tools in a structured way. When you connect Alpha Ladder to Claude, Cursor, or any MCP-compatible client, the client gains a set of **tools** — named functions it can invoke on your behalf. Alpha Ladder's server exposes these tools over **Streamable HTTP** at `POST https://www.alphaladder.hk/mcp`. Each time the AI decides it needs to look up an account balance or place a trade, it calls the appropriate tool, gets back a structured result, and incorporates it into its response — all without you leaving the conversation.

## Products and modules

Alpha Ladder exposes one product line via MCP:

### AgentX — financial workflows

AgentX is Alpha Ladder's financial operations layer for AI agents. It gives an agent structured access to account balances, fiat money movement, OTC FX conversion, transaction history, and wealth-product subscription flows. AgentX is **fiat-only**.

Its main workflows are:

- **Deposit** — fiat deposit flows, including bank instructions and submission steps.
- **Withdraw** — fiat withdrawal flows, including quote checks, destination selection, verification-code handling, and execution.
- **OTC exchange** — indicative rates, locked quotes, confirmation, and trade-detail lookup for FX conversion.
- **Wealth** — fixed-income products and subscription flows, including investor eligibility, product listings, agreements, and order submission.

Most AgentX flows are multi-step and stateful: the agent typically checks permissions, gathers quotes or destination details, asks for confirmation, and only then executes the final action. AgentX exposes **25 MCP tools** in total. For a full breakdown of accounts, currencies, and the first-party vs. third-party model, see [AgentX Concepts](/developers/financial-services/concepts).

## Read-only vs. destructive tools

Every MCP tool is annotated with a `readOnlyHint`. **Read-only tools** can be called freely by the AI as part of gathering information. **Destructive tools** — those that move money or submit transactions — carry `readOnlyHint: false` and **require explicit user confirmation** before the AI client will execute them.

The destructive tools are:

- `execute_fiat_withdrawal`
- `confirm_otc_trade`
- `fip_subscribe`
- `execute_third_party_fiat_deposit_submit`

When the AI reaches a step that calls one of these tools, it will pause and ask you to confirm before proceeding. This is enforced by the MCP protocol, not just convention.

## First-party vs. third-party (AgentX)

AgentX tools operate in one of two modes depending on how funds flow: **first-party** (directly through Alpha Ladder's own infrastructure) and **third-party** (routed via a custody partner, currently COBO). The distinction affects which currencies are supported, fee structures, and the shape of API responses.

For a full explanation of this model, account types, and how minor units work, see [AgentX Concepts](/developers/financial-services/concepts).