Where is the dashboard?
The core stack does not include one. VoicEra is API-first: http://localhost:8000/docs gives you an interactive console for every endpoint.
On http://localhost:3000, started with the rest of the stack. It covers agents, numbers, campaigns, knowledge documents, call history, and per-call latency. Its container runs the Next.js development server, so build it properly before exposing it. See Dashboard and Operating via the API.
Why is the database on port 27018?
The container listens on 27017; the host mapping is 27018 so it cannot collide with a MongoDB you already run locally. From your machine use 27018; inside the Compose network services use mongodb:27017.
It is FerretDB — the MongoDB wire protocol on top of PostgreSQL — not MongoDB. See Data store.
Do I need a GPU?
Not with cloud model providers. The core stack runs on 2 CPU cores and 4 GB of RAM.
A GPU is only needed to self-host models with the model server.
Can I use only OpenAI?
Yes. openai registers speech-to-text, text-to-speech, and a language model, so one credential covers all three. google and sarvam do the same.
Mixing is common — Deepgram for STT, Cartesia for TTS, OpenAI for the LLM. The choice is per agent.
Why are my provider dropdowns or catalogs empty?
Catalogs filter to providers you have stored credentials for. Store them first:
Why did my campaign pause itself?
The circuit breaker tripped — by default, more than 50% of calls failing within a 300-second window, over a minimum of 5 calls. It exists so a broken agent burns a handful of calls instead of the whole list.
Find out why the calls failed, fix it, then POST /campaign/{id}/resume. See Troubleshooting campaigns.
Why does my browser test call have no transcript?
Browser sessions do produce transcripts and recordings — the runtime registers a call_type: web call log on connect. If one is missing, check the runtime log for Registered web call call_id= and confirm MinIO is reachable.
Do I need a public URL?
Only for real phone calls. Your telephony provider fetches /answer over HTTPS and opens a WSS connection for audio — both inbound, so the runtime must be publicly reachable.
For evaluation, a websocket agent needs no public URL and no telephony account. See Public voice URLs.
My agents stopped answering after I changed a setting. Why?
Almost certainly VOICE_SERVER_BASE_URL. The answer URL is baked into the provider application when an agent is created, so changing it later does not update existing agents — your provider keeps calling the old address, and nothing reaches VoicEra to log.
PATCH each affected agent to re-provision, or recreate it.
How do I change a default password?
Edit .env and recreate the affected containers. Change MONGODB_PASSWORD, MINIO_ROOT_PASSWORD, and REDIS_PASSWORD at minimum.
Changing MONGODB_PASSWORD after the volume exists does not update the PostgreSQL user — the same credentials serve both layers. Set it before the first start, or change it inside Postgres too.
What happens if I lose PROVIDER_AUTH_ENCRYPTION_KEY?
Every stored provider credential becomes permanently undecryptable. There is no recovery path and no re-encryption tool — each organisation must re-enter every provider key.
Back it up with the same care as the database, and store it alongside your backups.
Is there a default login?
No. The first POST /users/signup creates the user, an organisation, and a super_admin membership. Whoever signs up first owns the deployment, so do it immediately after starting.
How many calls can run at once?
DEFAULT_ORG_CONCURRENCY_LIMIT caps simultaneous calls per organisation, default 10. Campaigns can set a lower max_concurrency.
In practice your telephony account’s channel limit or your model vendor’s rate limits usually bind first. See Call concurrency.
Can I scale the services?
The API, runtime, and ARQ worker scale horizontally. The runtime needs session affinity, since each live call holds one WebSocket.
The campaign orchestrator must run as exactly one replica. Its state is in-memory and it uses Redis pub/sub, which fans out to every subscriber — two replicas would dial each campaign at twice its configured rate.
Does VoicEra switch language mid-call?
No. An agent declares a primary language and optional secondary ones, but nothing switches during a call. Choose a provider whose model covers the languages you expect, or run separate numbers per language.
Where are recordings stored?
MinIO, under voicera-calls/{org_id}/{call_id}/. Fetch them through the authenticated API rather than the bucket:
Nothing expires automatically — set a retention policy.
How do I back up?
Three stores together: PostgreSQL (via pg_dump), MinIO, and the Chroma volume. Redis is ephemeral. Store PROVIDER_AUTH_ENCRYPTION_KEY with the backup — credentials are useless without it.
docker compose down -v deletes all four volumes at once, irreversibly.
Is there a CI pipeline?
No. There is no .github/ directory. Run the test suites yourself before opening a pull request — see Testing.