Skip to Content
cuelo.

06Ship

Ship to production

What changes between cuelo dev on your laptop and an agent taking real calls.

Moving from cuelo dev to production comes down to three things: lock it down, size it, and run it close to your providers.

Callerswss://TLS proxyLB · idle > callworker 1cuelo serve · modelsworker 2cuelo serve · modelsworker Ncuelo serve · modelsProvidersSTT · LLM · TTSclose region = low latencyPostgres / MySQLsessions · eventsVolume / bucketrecordings
calls stay on the worker their WebSocket landed on; state lives in the database.

The checklist

StepWhyPage
1Set serve.api_keysThe server refuses a public host without them, and they protect calls and recordings.Security
2Set serve.cors for browser appsBrowsers on other origins are blocked by default.Security
3Run cuelo machine checkReal-time capacity depends on the turn-taking stack and your CPU.Deploy
4Set serve.workersOne process uses about 4 cores, so add workers beyond that.Deploy
5Move the store to Postgres or MySQLSQLite is per machine.Storage
6Put TLS in front and allow long WebSocketsBrowsers need wss:// for the mic, and idle timeouts must exceed call length.Deploy
7Bake model weights into the imageThe first call shouldn't wait on a download.Deploy
8Deploy near your providersNetwork round trips are a large share of latency.Deploy

In this section

On this page