Skip to main content

​
What Is Function Calling?

Function calling lets your AI agent execute actions during a call. Instead of just talking, your agent can:
  • Book appointments in your calendar
  • Look up orders in your systems
  • Update CRM records
  • Send confirmations
  • End the call gracefully when the conversation is complete

Real-Time Action

The agent invokes tools mid-conversation and uses the results in its next reply.

Natural Experience

Callers don’t notice the tool call. It feels like talking to an agent who has your systems open in front of them.

No-Code or Your Own Server

Build tools visually in the Integrations platform, or connect your own MCP server through the API.

Explicit Opt-In

Only tools you allowlist are available to the agent. Nothing attaches by default.

​
Built-in Tools

TopCalls provides built-in tools that are automatically available during calls.

​
end_call

The end_call tool allows your AI agent to gracefully end the call when the conversation is complete.
The AI uses end_call automatically when appropriate. The agent decides when to end based on context: user says goodbye, all questions answered, or clear conversation conclusion.
When the AI uses end_call:
  • User says “goodbye”, “thanks, that’s all”, etc.
  • All questions have been answered
  • User explicitly asks to end the call
  • Conversation naturally concludes
What happens:
  1. AI decides to end the call
  2. AI speaks a contextual farewell message
  3. Call is terminated
You can influence end-call behavior in your instructions: “Always confirm the next steps before ending the call” or “Offer to help with anything else before saying goodbye.”

​
Custom Tools

There are two ways to give the agent custom tools.

​
Through the Integrations Platform (No Code)

Build a flow in the Integrations platform (for example “book a callback in Google Calendar” or “send an SMS confirmation”), then select it as a tool in the campaign editor. During calls, the agent can invoke the flow and use its result in the conversation. See Integrations for the full catalog of connectable tools.

​
Through Your Own MCP Server (API)

If you run your own tool server that speaks the Model Context Protocol, connect it per call: 
{
  "phone_number": "+14155551234",
  "instructions": "You are Rachel, an appointment coordinator...",
  "mcp_url": "https://tools.your-app.com/mcp/abc123",
  "mcp_token": "YOUR_MCP_BEARER_TOKEN",
  "mcp_tool_allowlist": ["check_availability", "book_appointment"],
  "tool_call_timeout_ms": 3000
}
FieldWhat it does
mcp_urlYour MCP server URL. The gateway connects at call start and merges your tools into the agent’s tool list.
mcp_tokenBearer token the gateway sends to your server. Redacted from logs.
mcp_tool_allowlistNames of the tools the agent may invoke. When absent or empty, no remote tools are attached.
tool_call_timeout_msPer-invocation timeout (500 to 10000 ms, default 3000).
The allowlist is an explicit opt-in. Listing your server with mcp_url alone attaches zero tools; name each tool the agent should have. This keeps the prompt small and the agent focused.

​
How a Tool Call Works

1

1. AI Decides

During the conversation, the AI decides a tool is needed based on what the caller asked.
2

2. Invocation

The gateway invokes the tool with the arguments the AI chose.
3

3. Result

The tool’s result is fed back to the AI.
4

4. Response

The AI works the result into its next reply and the conversation continues.
User: "Can you check if 3 PM tomorrow is available?"

AI: [Invokes check_availability with date and time]
    → Tool returns: {"available": true}

AI: "Yes, 3 PM tomorrow is available! Would you like to book it?"

​
Error Handling

If a tool call fails or times out, the AI is told and recovers conversationally:
“I’m having trouble reaching the calendar right now. Can I take your preferred time and have someone confirm it shortly?”
Design your tools to return clear error messages; the AI relays the situation naturally instead of exposing raw errors to the caller.

​
Best Practices

​
Do This

  • Clear names and descriptions: The AI picks tools based on their descriptions, so write them precisely
  • Validate inputs: Always validate arguments on your side before acting on them
  • Return structured data: Use consistent response formats
  • Fast responses: Return results quickly (under 2 seconds) so the conversation stays natural
  • Allowlist narrowly: Give the agent only the tools this call needs

​
Avoid This

  • Vague descriptions: “Helper function” doesn’t help the AI decide when to use it
  • Slow tools: Long tool calls create awkward silence on the line
  • Too many tools: Limit to 5-10 tools per call for best results
  • Secrets in prompts: Keep credentials on your server, never in instructions

​
Next Steps

Integrations

Build no-code tools and flows for your agent.

Making Calls

Learn how to include tools when making calls.