Handling Tool Calls

Monitor and respond to AI agent tool executions in real-time on the client side.

Overview

When AI agents use tools (functions) during conversations, you can monitor these tool calls on the client side. Tool calls and their results are now part of the conversation history, making them easy to display and track. This is useful for:

• Real-time UI updates: Show users what actions the AI is performing
• Progress indicators: Display loading states during tool execution
• Interactive confirmations: Allow users to approve/deny tool actions
• Analytics and logging: Track tool usage and effectiveness
• Conversation history: Tool calls appear in the conversation alongside messages

Basic Tool Call Monitoring

You can monitor tool calls in two ways:

1. Through Conversation History

Tool calls and results are automatically added to the conversation history:

import { Micdrop } from '@micdrop/client'

// Access conversation including tool calls and results
const conversation = Micdrop.conversation

conversation.forEach((item, index) => {
  switch (item.role) {
    case 'tool_call':
      console.log('Tool called:', item.toolName)
      console.log('Parameters:', JSON.parse(item.parameters))
      break
    case 'tool_result':
      console.log('Tool result:', item.toolName)
      console.log('Output:', JSON.parse(item.output))
      break
    case 'user':
    case 'assistant':
      console.log(`${item.role}: ${item.content}`)
      break
  }
})

2. Through Real-time Events

Listen for tool call events using the Micdrop client:

import { Micdrop } from '@micdrop/client'

// Listen for tool calls from the AI agent
Micdrop.on('ToolCall', (toolCall) => {
  console.log('AI is using tool:', toolCall.name)
  console.log('Parameters:', toolCall.parameters)
  console.log('Result:', toolCall.output)
})

Only tools with the emitOutput option enabled will emit ToolCall events.

Tool Call Structure

Each tool call event contains complete information about the tool execution:

interface ToolCall {
  name: string // Name of the tool that was called
  parameters: any // Parameters passed to the tool
  output: any // Result returned by the tool
}

React Example

Update your interface based on specific tool calls:

import { useMicdropState } from '@micdrop/react'
import { useState, useEffect } from 'react'

function SmartInterface() {
  const [userProfile, setUserProfile] = useState(null)
  const [emailStatus, setEmailStatus] = useState(null)
  const { isStarted } = useMicdropState()

  useEffect(() => {
    if (!isStarted) return

    const handleToolCall = (toolCall) => {
      // Handle different tool types
      switch (toolCall.name) {
        case 'save_user_data':
          // Update user profile in UI
          if (toolCall.output.success) {
            setUserProfile({
              ...toolCall.parameters,
              id: toolCall.output.userId,
              lastUpdated: new Date(),
            })
          }
          break

        case 'send_email':
          // Show email status
          setEmailStatus({
            sent: toolCall.output.sent,
            messageId: toolCall.output.messageId,
            recipient: toolCall.parameters.to,
            timestamp: new Date(),
          })
          break

        case 'search_database':
          // Could trigger a UI update with search results
          console.log(`Found ${toolCall.output.results.length} results`)
          break
      }
    }

    Micdrop.on('ToolCall', handleToolCall)

    return () => {
      Micdrop.off('ToolCall', handleToolCall)
    }
  }, [isStarted])

  return (
    <div className="smart-interface">
      {userProfile && (
        <div className="user-profile">
          <h3>User Profile</h3>
          <p>Name: {userProfile.name}</p>
          <p>Email: {userProfile.email}</p>
          <p>ID: {userProfile.id}</p>
          <small>
            Last updated: {userProfile.lastUpdated.toLocaleString()}
          </small>
        </div>
      )}

      {emailStatus && (
        <div className="email-status">
          <h3>Email Status</h3>
          <p>
            {emailStatus.sent ? '✅ Sent' : '❌ Failed'} to{' '}
            {emailStatus.recipient}
          </p>
          {emailStatus.messageId && (
            <small>Message ID: {emailStatus.messageId}</small>
          )}
        </div>
      )}
    </div>
  )
}

Server-side implementation

Agents can pass tool calls output to the client by emitting the ToolCall event.

For example, the OpenAI Agent emits the ToolCall event when a tool is called with the emitOutput option enabled.