Skip to main content
Console mode is the fastest proof the agent works. It runs the full pipeline in your terminal: no LiveKit, no backend, no database, no browser.
Talk into your microphone. The agent greets you first, then answers.

​
What it needs

Three keys in agent/.env: Nothing else. LIVEKIT_* can be blank, Postgres can be down, and the backend does not have to exist.
Console mode is not offline. It is a cascaded pipeline calling the real Deepgram, Cerebras and Inworld APIs. What it drops is the transport and the storage, not the providers.

​
What it skips

src/agent.py sets CONSOLE_MODE from sys.argv, and two things branch on it:
  • No participant wait. There is no room and no remote participant, so the worker starts the session immediately.
  • No call reporting. CallReporter.from_config(console_mode=True) stays silent whatever BACKEND_REPORTING_ENABLED says, so a console run never writes a row to the Calls page.
Everything else is the same code path the browser and the phone take: the same Assistant, the same persona file, the same models.

​
Where to go next

If console mode talks and the browser does not, the problem is in the transport or the config, not the agent. Start at Troubleshooting.