---
title: MCP
description: "Connects the agent to a Model Context Protocol server over HTTP, so it can use the tools and data that server exposes."
---

**MCP** connects the agent to a Model Context Protocol server. Whatever tools the
server exposes become available to the agent, with one configuration: the server's
URL and the headers it needs.

:::warning[Your server, your responsibility]
The form says it plainly: MCP servers are not provided or verified by Callab AI. It
is the client's or developer's responsibility to ensure the security and
reliability of the MCP server. Every tool the server exposes is a capability the
agent may use on a live call.
:::

![The MCP form.](/media/build/mcp-form.webp)

## Before you start

An MCP server reachable over HTTPS from the internet, and whatever token it expects
in a header. A server on your laptop or behind a VPN is not reachable.

## Configuring it

Open **Tools** under **Build**, choose **Add tool**, then **MCP**. After the
[shared fields](/build/tools/overview#fields-every-tool-shares):

| Field | Type | Default | Notes |
|---|---|---|---|
| **Tool Title** | Text | `MCP Server` | Name the server: `Inventory MCP`. Letters and spaces only; its ID, `mcp_server` by default, is shown under the field. |
| **Description** | Text | `Connect to Model Context Protocol server to extend agent capabilities` | Describe what the server offers so the agent knows when to reach for it: `Stock levels and delivery dates for any product`. |
| **AI Response** | Text | `Let me connect to the MCP server to access additional tools and data.` | Rewrite it; callers do not know what an MCP server is. |
| **Server URL** | URL | `https://your-mcp-server.com` | The server's HTTP endpoint. |
| **HTTP Headers** | JSON object | `{ "Content-Type": "application/json", "Authorization": "Bearer YOUR_TOKEN" }` | Sent with every request. Replace `YOUR_TOKEN`. |

Choose **Add Tool**, then **Save** the agent.

## Behaviour and limits

- The agent sees every tool the server lists. Expose only what the agent should be
  able to do; there is no per-tool switch on this side.
- Latency is the server's. A slow tool call is a pause on the call, so keep the
  server close to the caller and its tools quick.
- Each tool call to the server is recorded on the call's page with its request and
  response. See [Call details](/operate/call-details).
- One tool per server. Add several MCP tools for several servers.

## Troubleshooting

- **The agent says it will connect and nothing happens.** The URL is the
  placeholder or the server is unreachable from the internet. Test it with a
  request from outside your network.
- **401 or 403 from the server.** The **HTTP Headers** still contain `YOUR_TOKEN`,
  or the token lacks the scope the server requires.
- **The agent uses the wrong tool on the server.** Tool names and descriptions on
  the server side are what the agent reads. Make them specific there.

## Related

- [Tools overview](/build/tools/overview)
- [Live Webhook](/build/tools/live-webhook)