Ch 8: ElevenLabs Voice Agents

Shipping a voice agent is 40% code and 60% operations. This chapter is the operations half — six platforms, one Dockerfile, one CI pipeline, and a production checklist you can print and pin above your monitor.

Section 1

1. Railway — Fastest Path

Push repo to GitHub.
Railway → New Project → Deploy from GitHub → pick repo.
Variables tab → paste all keys from .env.
Set start command node dist/index.js. Deploy.

Claude Code — Terminal

Screenshot placeholder

Railway dashboard with a green Active deployment and a public URL.

Figure — Railway dashboard with a green Active deployment and a public URL.

Section 2

2. Render — Almost Identical

Same flow, slightly cheaper for always-on workers. Good WebSocket support.

Section 3

3. Docker — the Universal Package

dockerfileclaude-code

FROM node:20-alpine
WORKDIR /app
COPY package*.json ./
RUN npm ci --omit=dev
COPY . .
RUN npm run build
EXPOSE 3000
HEALTHCHECK CMD wget -qO- http://localhost:3000/health || exit 1
CMD ["node", "dist/index.js"]

bashclaude-code

docker build -t voice-agent .
docker run -p 3000:3000 --env-file .env voice-agent

Section 4

4. AWS — Production Scale

ECS Fargate for stateless containers.
ALB in front for TLS + WebSocket routing.
RDS Postgres for durable state.
Secrets Manager for keys.

Section 5

5. Cloudflare Workers

Edge functions with global latency. Great for the HTTP webhook layer. Cannot hold long-lived Twilio WebSockets — pair with a Durable Object or an origin server.

Section 6

6. Vercel

Perfect for the dashboard front-end. Serverless functions are OK for short jobs; use a dedicated worker (Railway) for long-lived voice sockets.

Section 7

7. Logging

tsclaude-code

import pino from "pino";
export const log = pino({ level: process.env.LOG_LEVEL ?? "info" });
log.info({ call_id, ms_total }, "turn.done");

Section 8

8. Monitoring

Uptime — BetterStack pinging /health every 30s.
Metrics — Grafana Cloud dashboards (latency p50/p95, cost/call).
Errors — Sentry with source maps.
Alerts — PagerDuty on p95 > 2s for 5 minutes.

Section 9

9. Secrets Management

Never commit .env. Ever.
Store secrets in the platform (Railway Variables, AWS SM, Cloudflare Secrets).
Rotate every 90 days. Automate rotation for Twilio + ElevenLabs.

Section 10

10. API Keys — Per Environment

Separate keys for dev, staging, prod. Never share across environments.

Section 11

11. Security

Verify every webhook signature (Twilio, Stripe, ElevenLabs).
Rate limit per IP + per phone number.
Strip PII from logs (mask phone middle 4 digits).
TLS only. HSTS on the dashboard domain.

Section 12

12. Rate Limits

tsclaude-code

import rateLimit from "express-rate-limit";
app.use("/api", rateLimit({ windowMs: 60_000, max: 60 }));   // 60 req/min/IP

Section 13

13. Scaling

Autoscale on CPU or on active WebSocket count.
Sticky sessions for calls (or hand off state via Redis on reconnect).
Warm pool of 2 instances always up to avoid cold-start on first ring.

Section 14

14. CI/CD

PR opens → run test, lint, typecheck.
Merge to main → build Docker image → push to registry.
Deploy to staging automatically. Prod behind a manual approval.
Smoke test after deploy: place one synthetic call, assert audio returned.

Section 15

15. GitHub Actions — Complete Pipeline

yamlclaude-code

name: deploy
on: { push: { branches: [main] } }
jobs:
  build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: 20 }
      - run: npm ci
      - run: npm test
      - run: npm run build

Section 16

16. Production Checklist

Health endpoint returning 200.
Structured logs shipping to a real log store.
Alerts on error rate, p95 latency, cost/day.
Secrets in platform vault, not repo.
Webhook signature verification.
Rate limits at the edge and app layer.
Docker image reproducible from any commit.
Rollback tested — one command back to last green build.
Synthetic call in CI on every deploy.

← Chapter 7 · Advanced Back to chapters Chapter 9 · Debugging →