Resources
FAQ
Common questions, answered with the limits included.
Product
What do I need to make a real phone call?
Four things: an agent, an account with a telephony provider (Twilio, Plivo or Exotel), a voice engine connected to speech and language models, and a wallet balance. Without them you can still build and test: see Testing without a phone.
Which languages are supported?
Whatever the speech models connected to your deployment support. The platform itself does not fix a list. The in-house models have been exercised in Hindi and English. See Models.
How fast does the agent answer?
The response time is the end-of-turn wait plus speech-to-text, the language model's first sentence and the first audio from text-to-speech. On the development setup it measured a little over one second. It is not a guaranteed figure: it depends on your models and network. See Voice engine.
Can the agent use a third-party speech or voice provider?
Yes, for some stages, with your own provider account. Add the key in Settings → Providers and pick the provider on the agent: Deepgram or Sarvam for speech-to-text (DEEPGRAM_API_KEY, SARVAM_API_KEY), ElevenLabs or Sarvam for the voice (ELEVENLABS_API_KEY, SARVAM_API_KEY), and OpenAI, Groq, OpenRouter or Azure OpenAI for the language model (OPENAI_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY, or AZURE_OPENAI_API_KEY with AZURE_OPENAI_ENDPOINT). The next call uses the key; nothing restarts. Without a key, or for a provider the catalogue lists with implemented: false, the call runs on our own models and the call's log says so in a line starting with note:. If the provider rejects the key mid-call, the rest of the call runs on our own models. The provider bills you directly for what you use with them. See Models.
How many calls can run at once?
Two limits apply: the workspace's concurrency limit from its plan, and the engine's capacity, which the operator sizes to the model servers (8 calls per engine process by default). Calls over the workspace limit wait in the queue.
Is there a library of ready-made agents?
The setup assistant has a gallery of templates across several industries that it turns into an agent on your account, after you approve its plan. It needs an AI backend configured by the operator. See Setup assistant.
Telephony
Do you provide phone numbers?
The API can search for, buy and release numbers through your provider account in live mode. In mock mode these are simulated. You can also bring a number you already own. See Inbound and numbers.
Has it carried real calls?
A small number of real calls on Exotel, during development. Twilio and Plivo support is implemented and tested against simulators, not yet on a real call. There has been no load test with real calls.
What happens when all lines are busy?
A call that reaches a full engine is refused without a spoken message. Configure a busy announcement at your telephony provider.
Data
Where are recordings and transcripts kept?
In the deployment's own storage and database. They are not encrypted at rest by the application; use disk encryption. Recording links are signed but do not expire. Recordings are deleted after a retention period, 90 days by default (RECORDING_RETENTION_DAYS; 0 keeps them for ever); Settings → Compliance shows the period in use. Recording can be turned off for the whole account there, or per agent on its Phone tab, which can also set an optional recording notice.
Is personal data removed from transcripts?
Card, account and identity numbers and one-time codes are masked by pattern, and keypad entries marked sensitive are silenced in the recording. Pattern matching misses things. Do not rely on it as your only control.
Are webhooks signed?
No. Set a secret header on the agent's webhook and check it on your side.
Money
What does it cost?
One per-minute rate per plan, phone line and our own models included: see Pricing for the current list. Calls are billed by the second from a prepaid wallet; the rates that apply to your workspace are in Settings → Billing and GET /billing/price-book. See Billing.
Can I pay online?
Top-ups go through a payment provider interface. A mock provider is used in development. Adapters for two payment providers exist but have only been tested against local fakes, so a deployment must run them in test mode before taking money.
Can a call take my balance below zero?
Yes. A call is not cut off when the balance runs out mid-call. A new call with no balance ends balance-low, and batches and campaigns pause until a top-up.
Compliance and security
Does the platform make my calls legal?
No. It gives you tools: calling hours, a do-not-call list, consent and daily caps, an AI disclosure line and redaction. You are responsible for following the laws that apply to your calls. No official do-not-disturb registry is integrated; the registry check uses a list you upload. See Compliance toolkit.
Has the platform been audited?
No external security audit or penetration test has been done, and there are no compliance certifications. See Security notes.
Is there single sign-on?
No. Logins use an email address and password, with an optional time-based second factor.
Operating it
Can it run highly available?
The described deployment is a single machine. There is no built-in high availability or zero-downtime upgrade. See Deployment.
Has the production stack been run?
Its configuration has been validated, but it has not yet been started on a real server.