**MCP** ([Model Context Protocol](https://modelcontextprotocol.io)) is an open standard for an AI
assistant (Claude Desktop, Claude Code, Cursor, or your own agent) to operate external tools in a
structured way, instead of through free text. **Aymar Agents** has its **own MCP server**: it exposes
your account — leads, contacts, conversations, appointments and tickets — as a set of *tools* your
assistant can read and, with your confirmation, write to.

## What it is

- A **deterministic adapter**, with no language model of its own: the LLM that decides which tool to
  call always lives in your client (Claude Desktop, Claude Code, Cursor…), never in this server.
- **Exactly what the public API allows** (`/public/v1`), not one endpoint more — its tools are generated
  by reading the same OpenAPI definition that documents the [REST reference](/en/api/getting-started/), so
  anything you can do over MCP you could also do with a direct `curl` call to the API. See the
  [full tools catalog](/en/mcp/tools/).
- **No access to your database, nor to any data the API doesn't already expose**: every tool call
  forwards your own API key to `/public/v1`, as-is — the MCP server never stores your key and has no
  other path into your account. All control (permissions, plan, limits, isolation per account) is
  enforced by the API, not by this adapter.

## What it isn't

**Don't confuse this with the "external MCP servers"** you can connect to your own AI agent from
**Settings → MCP servers** (BYO-MCP) — they run the opposite direction:

| | Aymar Agents's MCP server (this section) | External MCP servers (BYO-MCP) |
| --- | --- | --- |
| Direction | Exposes **your Aymar Agents account** to your external AI assistant | Plugs **third-party** tools into your account's own AI agent |
| Who connects it | You, from your MCP client (Claude Desktop, Claude Code, Cursor…) | You, from the Platform, so the agent that serves your customers can use it |
| Where it's configured | In your own MCP client's configuration (see [Connect your assistant](/en/mcp/connect/)) | In the Platform's **Settings → MCP servers** |

If what you're after is giving your AI agent access to external tools (your CRM, your catalog, your
calendar…), that's the other piece — see the **[Settings & integrations](/en/guides/ajustes-integraciones/)**
guide, section 8. This section is about the opposite: giving **your own AI assistant** access to
**Aymar Agents**.

## Requirements

- **Pro or Enterprise plan** — on the Starter plan, every tool responds with the same
  `403 PLAN_NOT_ELIGIBLE` the public API returns (see [Plans and limits](/en/api/plans-and-limits/)). The
  server always keeps listing the `aymar_status` tool so your client can show a clear message about why
  nothing else is available.
- **An API key** with the scopes for the resources you want to use — the same mechanism, and the same
  screen, as the key you'd use to call `/public/v1` directly (guide
  **[API keys](/en/guides/claves-api/)**, full scope catalog in
  **[Authentication](/en/api/authentication/)**). The tools your MCP client sees are exactly the ones your
  key can use — nothing more.

## How it works, in three steps

```
Your MCP client (Claude Desktop / Claude Code / Cursor / your own agent)
      │  Streamable HTTP, header X-API-Key: <your key>
      ▼
Aymar Agents's MCP server  (no database, no state, no LLM of its own)
      │  forwards your same key, call by call
      ▼
Aymar Agents's public API  (/public/v1 — everything is decided here: permissions, plan, limits)
```

Continue with **[Connect your assistant](/en/mcp/connect/)** for the actual configuration blocks. If you
also want your own backend to talk directly to the public API, with the same key, see **[My own AI and my
backend, talking to my data](/en/api/mi-ia-y-mi-backend/)**.