For AI agents: a documentation index is available at the root level at /llms.txt and /llms-full.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
Connect
BlogHelp CenterAPI PricingSign up
OverviewElevenCreativeElevenAgentsElevenAPIReception AIAPI referenceChangelog
OverviewElevenCreativeElevenAgentsElevenAPIReception AIAPI referenceChangelog
  • Get started
    • Overview
    • Quickstart
  • Configure
    • Overview
    • Voice & language
    • Knowledge base
    • Tools
    • Personalization
    • Authentication
  • Deploy
    • Overview
    • Environment variables
    • WhatsApp
    • Batch calls
  • Monitor
    • Overview
    • Users
    • Testing
    • Experiments
    • Versioning
    • Conversation Analysis
    • Analytics
    • Real-time monitoring
    • OpenTelemetry traces
    • Privacy
    • Cost optimization
    • CLI
  • Advanced
    • Events
    • Custom models
    • LLM cascading
    • Post-call webhooks
  • Resources
    • UI components
  • Guides
    • Chat Mode
    • Burst pricing
    • ElevenLabs' docs agent
    • Scaling user interviews
    • Simulate Conversations
LogoLogo
Login
Login
Connect
BlogHelp CenterAPI PricingSign up
On this page
  • Overview
  • WebSocket endpoint
  • Authentication
  • Example connection
  • Configuration
  • Control commands
  • Use cases
  • Limitations
  • Related resources
Monitor

Real-time monitoring

Observe live conversations and send control commands during active calls
Was this page helpful?
Previous

OpenTelemetry traces

Export OpenTelemetry traces to your observability stack as OTLP JSON.
Next
Built with

Real-time monitoring enables live observation of agent conversations via WebSocket and remote control of active calls. This feature provides real-time visibility into conversation events and allows intervention through control commands.

This is an enterprise-only feature.

Overview

Monitoring sessions stream conversation events in real-time, including transcripts, agent responses, and corrections. You can also send control commands to end calls, transfer to phone numbers, or enable human takeover during active chat conversations.

WebSocket endpoint

Connect to a live conversation using the monitoring endpoint:

wss://api.elevenlabs.io/v1/convai/conversations/{conversation_id}/monitor

Replace {conversation_id} with the ID of the conversation you want to monitor.

Authentication

Authentication requires:

  • API key permissions: Your API key must have ElevenLabs Agents Write scope
  • Workspace access: You must have EDITOR access to the agent’s workspace
  • Header format: Include your API key via the xi-api-key header

Example connection

1const ws = new WebSocket('wss://api.elevenlabs.io/v1/convai/conversations/conv_123/monitor', {
2 headers: {
3 'xi-api-key': 'your_api_key_here',
4 },
5});

Configuration

Before monitoring conversations, enable the feature in your agent’s settings:

1

Navigate to agent settings

Open your agent’s configuration page in the dashboard.

2

Enable monitoring

In the Advanced settings panel, toggle the “Monitoring” option.

Monitoring toggle in agent settings

3

Select events

Choose which events you want to monitor. See Client Events for a full list of available events.

The following events cannot be monitored: VAD scores, turn probability metrics, and pings.

The conversation must be active before you can connect to monitor it. You cannot monitor a conversation before it begins.

Control commands

Send JSON commands through the WebSocket to control the conversation:

End call

Terminate the active conversation immediately.

1// End the active conversation
2ws.send(JSON.stringify({
3 command_type: "end_call"
4}));
Transfer to phone number

Transfer the call to a specified phone number.

1// Transfer to a phone number
2ws.send(JSON.stringify({
3 command_type: "transfer_to_number",
4 parameters: {
5 phone_number: "+1234567890"
6 }
7}));
The transfer_to_number system tool must already be configured in the agent.
Realtime contextual update

Inject context or instructions into the active conversation so the agent can use the new information in its responses.

1// Send a contextual update to the agent
2ws.send(JSON.stringify({
3 command_type: "contextual_update",
4 parameters: {
5 contextual_update: "<your update text>"
6 }
7}));
Enable human takeover

Switch from AI agent to human operator mode for chat conversations.

1// Enable human takeover
2ws.send(JSON.stringify({
3 command_type: "enable_human_takeover"
4}));
Send message as human

Send a message to the user as a human operator in chat conversations.

1// Send a message as a human operator
2ws.send(JSON.stringify({
3 command_type: "send_human_message",
4 parameters: {
5 text: "How can I help you?"
6 }
7}));
Disable human takeover

Return control from human operator back to the AI agent.

1// Disable human takeover and return to AI
2ws.send(JSON.stringify({
3 command_type: "disable_human_takeover"
4}));

Use cases

Real-time monitoring enables several operational scenarios:

Quality assurance

Monitor agent conversations in real-time to ensure quality standards and identify training opportunities.

Human escalation

Detect conversations requiring human intervention and seamlessly take over from the AI agent.

Analytics dashboards

Build real-time monitoring dashboards that aggregate conversation metrics and performance indicators.

Call center oversight

Supervise multiple agent conversations simultaneously and intervene when necessary.

Automated intervention

Implement automated systems that analyze conversation content and trigger actions based on specific conditions.

Training and coaching

Use live conversations as training material and provide real-time feedback to improve agent performance.

Limitations

Asynchronous event delivery

Monitoring events are sent asynchronously to the conversation and may not arrive in the same order as the core conversation events. When processing events, do not rely on event order to reconstruct exact conversation timing.

Audio data not available

The monitoring endpoint streams only text events and metadata. Raw audio data is not included in monitoring events.

Historical event limit

Only approximately the last 100 events are cached and available when connecting to an active conversation. Earlier events cannot be retrieved.

Event filtering restrictions

VAD scores, turn probability metrics, and ping events cannot be monitored when custom event selection is enabled.

Connection timing

You must connect after the conversation has started. The monitoring endpoint cannot be used before conversation initiation.

Permissions required

API keys must have ElevenLabs Agents Write scope, and you must have EDITOR workspace access to monitor conversations.

Related resources

Post-call Webhooks

Receive conversation data and analysis after calls complete.

Agent Analysis

Configure success evaluation and data collection for conversations.

Client Events

Understand events received during conversational applications.

WebSocket API

Learn about the WebSocket API for real-time conversations.