gpt-live-demo

GPT-Live Voice Operator Demo

Overview

This skill packages a complete GPT-Live WebRTC voice demo:

Web frontend（scripts/index.html): a pure browser WebRTC client, including call controls, event log, and per-second billing display
Relay server（scripts/server.mjs): Node.js/Express, responsible for requesting a WebRTC session from OpenAI and forwarding delegated tasks to the Starchild agent
Deployment instructions（references/deploy.md): local development and production deployment steps

Architecture

浏览器 (index.html)
  │  WebRTC audio
  │  DataChannel (oai-events)
  ▼
OpenAI GPT-Live (gpt-live-1)
  │  client delegation 事件
  ▼
server.mjs (/api/agent)
  │  HTTP SSE stream
  ▼
Starchild agent (localhost:8000/chat/stream)
  │  结果回传
  ▼
GPT-Live → 语音播报

Quick start

1. Install dependencies

cd <skill-dir>/scripts
npm install

2. Prepare an OpenAI API key (required reading for installers)

This skill depends on OpenAI's gpt-live-1 realtime voice model,has no built-in key, and cannot apply for one on your behalf. Before use, please prepare it yourself:

Go to https://platform.openai.com/api-keys to create an API key (sk-... format);
The account must already have gpt-live-1 (Realtime Beta) access—apply for it in the OpenAI backend; if not enabled, it will report 403/model_not_found;
Write the key into an environment variable (or the .env）：

export OPENAI_API_KEY=sk-...   # 必填，需 gpt-live-1 访问权限
export PORT=3000               # 可选，默认 3000

When there is no valid key,/api/session will return 401, and the browser side will show "unable to start call".

3. Start the service

node server.mjs

Then open http://localhost:3000 to start the voice call.

Prerequisite: the Starchild agent's /chat/stream endpoint must be running at localhost:8000 ; if you use this skill within the Starchild platform, this endpoint is already available automatically.

File description

File	Description
`scripts/server.mjs`	Express relay server (WebRTC session proxy + brain bridge)
`scripts/index.html`	Browser frontend (WebRTC + UI + billing meter)
`scripts/package.json`	Node.js dependencies (express, openai)
`references/deploy.md`	Complete guide for local development and production deployment
`references/api-notes.md`	GPT-Live delegation API key points and known pitfalls

Scenarios for using this skill

When the user requests:

Start / demo the GPT-Live voice demo
Modify the voice operator's system prompt (LIVE_PROMPT）
Adjust the delegation strategy (when to delegate to the Starchild brain)
Deploy to production (Fly.io / Docker / Nginx reverse proxy)

→ read references/deploy.md for complete steps.

Notes

You must provide your own OPENAI_API_KEYbefore use (see "Quick start · Step 2"); gpt-live-1 (Realtime Beta) access must be enabled, otherwise the call cannot be established
The service listens only locally by default; before production deployment, add authentication in /api/session Add authentication
Tasks and voice history are persisted to the data/ directory (debounced writes to disk), and are not lost on restart; completed tasks are retained for 1 hour
Billing estimates are based on $0.05/minute (gpt-live-1 voice); backend brain call costs are billed separately