Skip to main content
LiveKit monitoring sends conversation evidence from your worker through the Egma SDK. Retell monitoring imports completed calls with the Retell API. You do not need a test suite or a simulation run for either setup.

​
Retell

Retell monitoring imports calls from Retell voice agents. It does not import text-only Retell conversations.

​
Connect the agent

Complete CLI sign-in and check egma/config.yaml for the agent. If it is not registered, run:
Set EGMA_AGENT_ID to that Egma agent ID. If you have already added a Retell connection, use its stored provider ID and key:
For a monitoring-only setup, load the Retell key into EGMA_RETELL_API_KEY. Discover the provider ID, set RETELL_AGENT_ID to the selected ID, then enable monitoring:
You can supply the key through standard input with --credentials-stdin instead. It accepts a JSON object with an apiKey field. Setup stores the Retell binding and starts importing calls. It does not create a simulation connection. One Egma agent monitors one Retell agent; reuse the existing binding instead of registering the same Retell agent again.

​
Verify an imported call

The first setup imports available completed calls from the previous 30 days. Egma checks for new completed calls about every 30 seconds while monitoring is enabled. Allow time for Retell to finalize a call and for the import to finish. Open Monitoring in your Egma project. Find a recent call from the agent, open its trace, and check the transcript against that call. If nothing appears:
  • Confirm that the call belongs to the Retell agent ID you selected.
  • Check that the call has ended and is available in Retell’s history.
  • Open Agents, select the agent, and check that Production monitoring is Active. Its last-received time shows whether a call has arrived.
  • If setup failed, resolve the authorization or provider error that the CLI returned before retrying.
To stop future imports, run:
Previously imported conversations stay available in Egma.

​
LiveKit

​
Create an Egma project key

In your initialized agent repository, run:
Copy the key when the CLI prints it. The CLI shows it once and does not save it. You can also create a key in Settings → API keys. Add these environment variables to the worker through your normal deployment secret store: The worker must be able to reach EGMA_URL. localhost refers to the worker’s own machine or container, so use a reachable address for a remote deployment.

​
Add the monitoring hook

Install the LiveKit Python SDK:
Add this import to the file that defines your job entrypoint:
Make this the first statement inside the entrypoint, before ctx.connect or AgentSession.start:
This requires Python 3.11 or newer and livekit-agents>=1.6.6,<1.9.
Apply the code and environment variables to the worker that handles real conversations, then restart or redeploy that worker. The hook adds Egma export alongside a compatible existing OpenTelemetry setup, including LiveKit Cloud observability. It sends spans in batches and flushes the final batch when the job ends. The hook ignores rooms whose names start with egma-sim-, so your tests do not also appear as production conversations. The separate simulation hook sends those rooms’ evidence to their simulation records. Use both hooks when the same worker handles testing and production; both use the same Egma settings. Keep the egma-sim- prefix reserved when naming your own production rooms.

​
Verify a conversation

Complete one normal conversation with the updated worker, then open Monitoring in the project that owns the API key. Open the new trace and check its transcript and timing. There is no LiveKit monitoring switch to enable in Egma. The first received trace confirms that the worker can send evidence to Egma. If no conversation appears, check the worker’s OpenTelemetry logs, the project key, and network access to EGMA_URL. To stop this worker from sending new production evidence, remove its monitor call and restart the worker. Existing conversations stay in Egma. Keep simulation if the worker still handles Egma tests.

​
Read transcripts and grades

The Monitoring table shows the agent, time, duration, p90 turn latency, and trace ID when those values are available. Select a trace to open its details and inspect the transcript, metrics, and grades. To grade incoming conversations, configure graders with production scope and the sample percentage you want. A missing grade can mean the conversation was outside that scope or sample, is still processing, or could not be graded. Check its grading state before treating it as a pass. Each grade has its own score and threshold. A combined score summarizes the selected grades; it does not decide an overall pass or fail. See Scores and thresholds.