What Is MCP?

MCP is a protocol that lets any compatible server advertise a list of tools — named, documented functions with typed inputs — and lets the AI Agent call those tools during a conversation. Because it is a standard, you do not write custom integrations for every system. If a vendor or an internal team publishes an MCP server, the agent can use it.

When you connect an MCP server to an agent:

1. At the start of every state change, the agent asks the server for its tool list.
2. Each tool becomes available to the agent for that state in the call, alongside its built-in tools.
3. If the agent decides a tool is the right answer, it calls the server, waits for the result, and uses it in the conversation.

When To Use an MCP Server

Use an MCP server when you need the agent to do something live against another system:

• CRM lookups. "Is this caller already a customer? What did they buy last?"
• Order or ticket status. "Where is order 4821 right now?"
• Inventory or availability. "Do we have that part in stock at the downtown branch?"
• Custom bookings. If your scheduling system is not one of the built-in integrations, expose it as an MCP server.
• Internal workflows. Kick off an approval, create a ticket, send a Slack message.

If you only need to act after the call ends — log it to a spreadsheet, create a CRM contact — use the webhook post-call action or Zapier instead. Those are simpler.

Before You Start

You will need:

• The MCP server's URL — usually an HTTPS endpoint ending in /sse or /mcp.
• An authentication token if the server requires one (most do — bearer tokens are standard).
• A short list of what the server provides. Ideally the server's own documentation, so you know what the tools do.
  

Adding an MCP Server to an Agent

1. Open the agent and go to the MCP Servers tab found in the Actions Tab.
2. Click Add Server.
3. Fill in the fields:
   • Name — a short label, e.g. "HubSpot CRM."
   • Description — what the server does, e.g. "Customer lookups and ticket creation."
   • URL — the server endpoint.
   • Authentication — pick Bearer Token (most common) and paste the token, or None for public servers.
4. In the Prompt Instructions field, write a short guide for the agent on when to use these tools. Example: "Use the CRM tools to look up customer records when the caller gives their name or phone number." The agent reads this when deciding whether to call a tool.
5. Enable the server and save.

Testing the Connection

1. After saving, place a test call that would naturally use one of the server's tools. If you added a CRM server, call in and say "I am calling about order 123."
2. Watch the call transcript or recording. You should see the agent call the tool and incorporate the result.
3. If the agent ignores the tool, sharpen the prompt instructions. If the tool errors, check the URL and token.

Comfort Messages for MCP Calls

External calls can take a second or two. Pair every MCP server with a comfort message so the caller does not hear silence while a tool runs. See the Comfort Messages for Tool Latency article. A good filler for a CRM lookup is "Let me pull up your account."

Security Notes

• Tokens are stored encrypted. They are sent only when the agent connects to the server, never to the caller.
• Only enable servers you trust. A tool can do anything the server allows it to do.
• Rotate tokens regularly on the server side and update them here.
• If an MCP server is misbehaving or compromised, disable it with the toggle — the agent stops calling it on the next new call.
  

1. Data Exposure: Do not expose sensitive information to agents that will interact with the public. Carefully review what data your MCP servers can access.
2. Write Operations: Exercise caution with MCP tools that can create, update, or modify data. Consider read-only configurations for public agents.
3. Delete Operations: Never allow public agents to delete information via MCP. Disable any delete tools in your MCP server configuration.
4. Tool Permissions: Review and limit the tools available to each MCP server. Use tool filters to restrict what actions the agent can perform.
5. Authentication: Ensure MCP servers use proper authentication. Avoid exposing servers with elevated privileges to public agents.

Troubleshooting

• Agent never calls the tool. The prompt instructions are probably too vague. Name the tool and the trigger — "When the caller asks about an order, use get_order_status."
• Tool returns an error. Check the server logs on your side. Most errors are auth (bad token) or permissions (the token does not cover the operation).
• Tool is slow. Add a comfort message. If it is consistently longer than 5 seconds, the caller will notice no matter what — work with the server team to speed it up.
• Tool list is empty on connect. The URL is likely wrong, or the server requires a different auth type.