Skip to main content

​
With a coding agent

Open the folder in Claude Code, then type /setup at its prompt. It asks one question, what this agent should do, and does the rest: copies .env.example to .env, runs scripts/doctor.py, names the variables still missing and where each comes from, rewrites agent/prompts/instructions.md for what you described, brings the stack up, and runs the doctor again. Agents other than Claude Code read AGENTS.md, which carries the same facts. Open the repo in one and say “set this up”. It will ask you for a LiveKit project (URL, API key, API secret) and three provider keys: Deepgram, Cerebras and Inworld. A free LiveKit Cloud project works.
The agent never handles your API keys. You paste them into .env yourself.Everything typed into a coding-agent conversation is written to that agent’s transcript on disk, so a key that passes through the chat is a key sitting in a file you did not know to look in. /setup names the variable and the file, then waits for you. That is the reason to trust this flow, not the price of it.
When the doctor passes, open http://localhost:5173, go to Agents, open the agent, then Test call and Start test call. Allow the microphone, then talk.

​
By hand

The six are LIVEKIT_URL, LIVEKIT_API_KEY, LIVEKIT_API_SECRET, DEEPGRAM_API_KEY, CEREBRAS_API_KEY and INWORLD_API_KEY. Everything else has a working default. LiveKit is external: the worker connects out to LIVEKIT_URL. The backend brings the schema to head on startup, so there is no migration step. Then http://localhost:5173, Agents, open the agent, Test call, Start test call. The Agents list has no such control, the agent’s own page does.
docker compose reads the root .env and nothing else. The agent/.env, backend/.env and frontend/.env files are for running one service by hand. A value written into the wrong one produces a config that works one way and not the other.
The console is http://localhost:5173, the backend http://localhost:8000 with OpenAPI at /docs. Both publish on 127.0.0.1, so nothing is reachable from the network.

​
Console mode, before anything else is up

The whole speech to model to speech loop in a terminal. No LiveKit, no backend, no database, no browser.
The agent greets you first, then answers. It needs DEEPGRAM_API_KEY, CEREBRAS_API_KEY and INWORLD_API_KEY in agent/.env and nothing else. It is not offline: it calls the real Deepgram, Cerebras and Inworld APIs, and drops the transport and the storage, not the providers. A console run never reports a call, so the Calls page stays empty whatever BACKEND_REPORTING_ENABLED says. If console mode talks and the browser does not, the problem is transport or config, not the agent.

​
Running the three services separately

Three terminals, one per service, each with its own .env.example.

​
If nothing speaks

Run the doctor rather than the logs. It names the cause and the fix, and Troubleshooting lists every symptom.