MCP Servers

What is an MCP server?

The Model Context Protocol (MCP) is an open standard for exposing tools to AI models. An MCP server is simply a remote HTTP endpoint that speaks this protocol. You don't install anything — you paste the server's URL (and optional authentication headers), and the assistant can call its tools while talking to your customer.

Mid-call tools vs. MCP servers

Both live under the Tools page. Use whichever fits:

A single custom HTTP endpoint you define by hand (URL, method, parameters).
Best for one-off calls to your own API. See [Custom Mid-Call Tools](/ai-assistants/custom-tools).

A remote server that exposes **many** tools at once, discovered
automatically. Best for connecting to a platform or a shared integration.

Connecting a server

Go to **Tools** in the sidebar and switch to the **MCP servers** tab, then click **New MCP server**.

- **Display name**: a friendly name shown when assigning the server (e.g. `HubSpot CRM`)
- **Identifier**: an internal id, lowercase letters and underscores only (e.g. `hubspot`)
- **Description**: optional, for your own reference
- **Server URL**: the remote MCP endpoint
- **Timeout**: how long to wait when connecting to the server

Leave this on **Auto-detect** unless you know otherwise. Auto-detect picks the
right transport from the URL:

```yaml
URL ends with /mcp  →  Streamable HTTP
URL ends with /sse  →  SSE (Server-Sent Events)
```

If the server requires a token, add it as a header. Headers are **stored encrypted**.

```yaml
Authorization: Bearer your_token_here
```

When you save, the server is tested automatically and a notification lists the tools it exposes.

Try it for free

Connection status & discovered tools

Each server is shown as a card with a live connection status.

The server is reachable. Its available tools are listed on the card.

The server couldn't be reached. The error is shown so you can fix the URL or headers.

The server hasn't been checked yet.

Use Test connection at any time to re-check a server and refresh its tool list.

Choosing which tools are exposed

By default, all of a server's tools are available to your assistants. Open Manage tools (on the server's edit page) to view each tool's description and parameters, and enable or disable individual tools.

Assigning a server to an assistant

Edit an assistant and go to the **Prompt & Tools** step.

In the **MCP servers** section, select one or more servers. The list shows
each server's connection status and tool count.

The assistant can now use the server's enabled tools during voice calls, the
web widget, and chat.

Authentication

The integration sends static headers with each request. It does not perform an interactive OAuth login.

Real-world examples

```yaml
Display name: DeepWiki
Identifier: deepwiki
Server URL: https://mcp.deepwiki.com/mcp
Headers: (none)
Tools: read_wiki_structure, read_wiki_contents, ask_question
```

The assistant can answer questions about any public GitHub repository.

```yaml
Display name: HubSpot CRM
Identifier: hubspot
Server URL: https://your-mcp-host.example.com/mcp
Headers:
  Authorization: Bearer pat-xxxxxxxx
```

The assistant can look up contacts, create deals, or update records during a call.

Best practices & security

• Keep tokens scoped to the minimum permissions the assistant needs.
• Use Manage tools to expose only the tools an assistant actually requires.
• Give each tool a clear name and description on the server so the AI knows when to use it.
• Mention the available capabilities in your assistant's system prompt so it uses them at the right moments.