Website demo

turncall-demo is a small server for a “talk to an agent” button on a website. It runs the real voice pipeline — the agent’s own STT, LLM and TTS, Silero VAD, Smart Turn — over WebRTC, with no database, Redis, project or carrier. The agents come from a file and the provider keys from the environment, so it runs where the platform cannot: a container that starts per visitor and is thrown away after the call. It authenticates nothing. Put it behind something that decides who may call — a CAPTCHA, a rate limit, a daily budget — because every minute is your provider bill.

Run it

It is the TurnCall image with another command:

docker run --rm -p 8090:8090 \
  -e DEMO_AGENTS_FILE=/demo/agents.json -v "$PWD/agents.json:/demo/agents.json:ro" \
  -e OPENAI_API_KEY=... -e DEEPGRAM_API_KEY=... \
  ghcr.io/kobikis/turncall:main turncall-demo

The agents file maps a name to an agent config, in the shape POST /v1/agents takes as config (Agents). The first is the default:

{
  "front-desk": {
    "system_prompt": "You are the receptionist at Northside Dental...",
    "first_message": "Thanks for calling Northside Dental. How can I help?",
    "llm": {"provider": "openai", "max_tokens": 200},
    "tts": {"provider": "deepgram", "voice": "aura-2-helena-en"},
    "tools": [{"name": "end_call", "description": "End the call after saying goodbye.", "parameters_schema": {}}],
    "analysis": {"enabled": false}
  },
  "restaurant": { "...": "..." }
}

A demo call has nothing to write to, so an agent may use only what needs no platform. The server refuses to start — naming the agent and the field — on an agent with:

a tool other than end_call (which says goodbye and hangs up), or a webhook tool
MCP servers, knowledge bases, ambience or a video avatar

An unknown field is refused too, as on the API. Every agent’s max_call_duration_seconds is capped at DEMO_MAX_CALL_SECONDS.

Settings

Variable	Default
`DEMO_AGENTS_FILE`	`demo-agents.json`	The agents, by name
`DEMO_MAX_CALL_SECONDS`	`180`	Ceiling on every agent’s call length
`DEMO_MAX_CALLS`	`1`	Calls one process holds at once; the next offer is a `429`
`DEMO_EXIT_AFTER_CALL`	`false`	Exit once the last call ends, so a per-visitor container stops billing
`WEBRTC_ICE_SERVERS`	public STUN	A TURN relay, when the server has no public address. See WebRTC
`HOST` / `PORT`	`0.0.0.0` / `8090`	Where it listens

Provider keys are the platform’s usual variables (OPENAI_API_KEY, DEEPGRAM_API_KEY, …). Without one an agent needs, the call is refused with a 503 naming it.

Endpoints

The browser speaks to it with pipecat’s SmallWebRTCTransport, as it does to /v1/webrtc/connect. The agent is named in requestData:

await client.connect({
  webrtcRequestParams: { endpoint: "/connect", requestData: { demo: "front-desk" } },
});


`POST /connect`	The SDP offer; answered with the SDP answer. `404` for an unknown agent, `429` when the line is busy, `503` while shutting down
`PATCH /connect`	ICE candidates for an open connection
`GET /agents`	`{"agents": [...], "default": "..."}`
`GET /ready`	`200` when a call can start, `503` (`busy` or `draining`) when not
`GET /live`	`200` while the process runs

On SIGTERM it stops taking calls and lets the one on the line finish, as the API does (graceful shutdown).

What it does not do

No call is stored: no transcript, recording, events, webhooks or post-call analysis. Nothing reaches a project, so project credentials, limits and suppression lists do not apply — the server’s own DEMO_MAX_CALLS and whatever fronts it are the only limits.

Getting Started

Guides

Run it

Settings

Endpoints

What it does not do

​Run it

​Settings

​Endpoints

​What it does not do

Run it

Settings

Endpoints

What it does not do