LiveAvatar Demos

Four curated demos, each a real open-source repo. This skill does every mechanical step: clone, install, provision, configure, launch.

What this skill cannot do — say this up front

You cannot verify that the avatar talks. Every demo ends at a microphone and a human ear. Do not claim success you can't observe, and do not stall waiting for something you can't reach. Name these moments as you hit them:

Everything else is yours. Do it without narrating each command.

Step 1: Get the LiveAvatar API key — before anything else

All four demos need it, and having it early removes most branching. Ask for it now, in your first message, alongside the catalog:

Grab an API key from https://app.liveavatar.com/developers and paste it here. I'll take it from there.

Do not proceed past Step 3 without it. You may clone and install while waiting, so the user is one paste away from a running demo — say that's what you're doing.

Never write a placeholder or fake value into an env file. Every setup script and entrypoint treats whatever is in the file as authoritative, and a bogus key produces a 401 that looks like an account problem.

Step 2: Pick the Demo

Show this table unless the user already named one.

Routing: fastest thing that talks → 1. Wants to see a voice pipeline, has LiveKit Cloud → 2. Already runs LiveKit rooms in production → 3. Wants speech-to-speech, GPT-Live, visuals synced to the avatar, or a starting point they intend to reshape into their own demo → 4; the reference picks the closest branch. Do not gate 4 on the user confirming access to a specific GPT-Live model — the repo's .env.example tracks the current release. "Just show me anything" → 1.

Demos 2 and 3 are the same repo, two entrypoints. Wanting both means cloning once.

Costs — state this once, before launching

• LiveAvatar sandbox sessions are free and duration-capped at roughly a minute. Demos 2 and 3 default to sandbox; demo 1 uses the shared public demo avatar.
• Demo 4 has no sandbox — production LiveAvatar billing from the moment the session starts, capped only by the account's default session limit; the repo sets no cap of its own. It also bills OpenAI GPT-Live audio tokens per turn.
• Demos 2 and 3 bill STT/LLM/TTS to the user's LiveKit Cloud project — including demo 2, where LiveAvatar owns the room. Room ownership and inference billing are separate. People assume otherwise. Demo 3 also bills RTC minutes.
• Demo 1 bills Anthropic tokens per turn, and only once the brain is wired.

Step 3: Preflight

Run only the checks for the chosen demo. Report anything missing with its install command; don't install runtimes without asking.

Demo 1:

node --version                 # need >= 22 (engines.node enforces it)
git --version
lsof -ti tcp:3003              # MUST be empty — see below

Port 3003 is pinned in both dev and start. An occupied port is fatal, not a warning, and there is no auto-increment. If lsof returns a PID, either free it or plan to bypass npm with npx next dev --port <free> — and remember the handoff URL changes with it.

Demos 2 and 3:

git --version
uv --version                   # install: brew install uv

Do not gate on python3 --version. uv venv reads requires-python from pyproject.toml and selects a compatible interpreter itself — a system Python 3.9 is irrelevant. Checking it produces a false negative that aborts a setup which would have worked. If uv is genuinely unavailable, fall back to python3.13 -m venv .venv && .venv/bin/pip install -e . and say you did.

Demo 3 also needs lk: brew install livekit-cli (tap livekit/livekit first if it doesn't resolve), then lk cloud auth — interactive, hand it to the user.

Demo 4:

node --version                 # need >= 20.12
pnpm --version                 # install: corepack enable, or brew install pnpm
git --version

Ports 8787 (server) and 5173 (Vite) — only 8787 matters; Vite relocates itself, the server does not. The reference covers moving it.

Step 4: Remaining Credentials

Ask for these in one message, after the demo is chosen — not before, since what's required depends on it.

For demo 1, read the Local-vs-Full section of its reference and pick a level before asking — asking for an Anthropic key the local run never uses blocks users who have only a LiveAvatar key.

Avatar ID: never ask. Every repo ships a working default in .env.example. Leave it.

Optional and demo-1-only, skip unless raised: NOTION_TOKEN + NOTION_DATABASE_ID, SLACK_WEBHOOK_URL.

Where to clone

Default to ../liveavatar-demos/<repo-name> relative to the current directory (demo 4: <repo-name>-<branch>, since every branch shares one repo name), and confirm the path before running git clone. Never clone inside the user's project tree — if cwd is their project, go up and out. If the target exists and is non-empty, stop and ask.

Step 5: Set Up and Run

Clone first, then read the repo's own agent docs before running anything: AGENTS.md, then CLAUDE.md, then docs/ and the README's troubleshooting section — whichever exist. The repo's docs are newer than this skill and win on commands and file layout; the reference below holds what the repo can't know — account-level gotchas, billing, and routing. If they contradict each other, follow the repo and note the drift.

Rules for every demo:

Secrets go in the env file the repo designates — never a committed file. Demos 1–3 use .env.local; demo 4 uses .env at the repo root (gitignored there). Both patterns ship an .env.example.

Write keys with an editor tool, not a shell command. Shell history and logs persist. The exception is a setup script that reads keys from the environment — the references cover those cases.

Copying .env.example is not enough. These files ship keys with empty values that are not equivalent to unset. The references name each one. Getting this wrong is the single most common way these demos die.

Never activate a venv. source .venv/bin/activate does not persist between your tool calls. Use uv run <command> every time.

Launch long-lived processes in the background and read the log rather than blocking on a foreground process. Interactive setup scripts are the exception — see the reference.

Report the ready signal you actually saw, then stop. Each reference names one. Every ready signal proves the process started, not that the avatar works. Do not extrapolate.

Step 6: Hand Off

Give exactly this, then stop:

1. What the user must do now, concretely. Demo 1: open http://localhost:3003, fill in a name and work email, click Chat with Wayne. Demos 2/3: a browser tab opened automatically — allow the microphone, then speak. Demo 4: open http://localhost:5173, click Start, allow the microphone, say "ready" when asked.
2. What you could not verify — that the avatar appears, speaks, and lip-syncs (demo 4: and that the overlays land). Ask them to confirm.
3. One customization pointer. Demo 1: prompt parts in src/lib/ai-sales/brain/prompt-parts/. Demos 2/3: instructions= in src/agent.py. Demo 4: the persona in server/prompts/*.md — and offer to go further; the repo's docs/ walk through adding visuals and repurposing the whole domain, and you can do that work.
4. How to stop, and what's billing. The local dev server costs nothing. The billable thing is the avatar session — it ends when the browser tab closes or the process dies. Say which one applies, and for demo 4 that sessions are production-billed, not sandbox.

Don't tour the codebase. They'll ask.

When Things Break

Each reference has a gotchas section with the specific failure, and demo repos with real agent docs put their troubleshooting there — check both. General triage:

1. Crash before any network call → an empty-but-present env var. Check LOG_LEVEL (demos 2/3) first; it's the most common and its traceback names nothing relevant.
2. Avatar connects but is silent → the speech leg, not the video leg. Demo 1: missing context (opening_text) or an LLM configuration pointed at an unreachable URL — silent failure, no error thrown. Demo 4: the GPT-Live leg failed to authenticate — GPT_LIVE_DEBUG=1 says why; the reference says what not to change.
3. Crash on the user's first sentence (demos 2/3) → model weights never downloaded.
4. 401 at session mint → bad or missing LIVEAVATAR_API_KEY. This error is clear and quotes the request body; trust it.
5. Session rejected at mint → max_session_duration above the account tier's cap, or a production avatar ID in sandbox mode.
6. Nothing in the browser → check the browser console before the server log. Usually mic permission.

If the symptom isn't in the reference's table, debug from the log, not the README.

Step 7: Collect Feedback

Trigger once the demo is handed off, or if the user abandons after 3+ failed attempts at the same step. If you're blocked on credentials and nothing is running, skip it.

Follow the liveavatar-feedback skill if it's available; if it isn't, skip silently rather than improvising a feedback channel. One line, and don't push if the user ignores it.

What to consult

• references/sales-agent.md — AI Sales Agent: interactive setup script, local vs. deployed, Vercel, env reference
• references/livekit-agent.md — both LiveKit demos: common setup, hosted flow, BYO flow, worker deploy
• references/gpt-live-demos.md — GPT-Live × Hyperframes: which branches exist, which repo docs to read, agent-safe setup, production billing, iteration tiers