FAQs
Quick answers to the questions customers and developers ask most often about Xebit.
Showing instructions for Xebit VoiceCloud.Showing instructions for Xebit VoiceKit, on your own infrastructure. Switch with the selector at the top. Updated 22 October 2026.
Short answers to the questions we hear most from teams evaluating, building on and running Xebit. Where the answer depends on whether you use Xebit VoiceCloud or Xebit VoiceKit, use the switch at the top of the docs to see the version that applies to you. For anything not covered here, the Platform Guide and the API Reference go into full detail.
Getting started
What is Xebit?
Xebit Corp. is a Delaware C-Corp with three AI products: Xebit Training Agent, Xebit VoiceCloud and Xebit VoiceKit. These docs cover the voice platform, Xebit VoiceCloud and VoiceKit: a voice AI platform for building, running and observing voice agents that handle real phone calls at volume. It covers the whole stack: the Agent SDK, the audio bridge, the agent server, the speech pipeline, orchestration, scheduling and campaigns, observability, the control plane and a SIP gateway that connects to your own carrier. You can use it as Xebit VoiceCloud, which we run, or as Xebit VoiceKit on your own infrastructure.
Should I start with Xebit VoiceCloud or Xebit VoiceKit?
Start with Xebit VoiceCloud if you want to go live quickly without running infrastructure, and your audio and transcripts can be processed outside your network. Choose Xebit VoiceKit if data has to stay inside your environment, if you need an air gapped deployment, or if you expect more than roughly 10,000 connected minutes a month, where owning the licence costs less. Deployment options compares the two in detail.
How do I get my first agent live?
Create an account and build the agent: the conversation, the business rules and what should happen when the call ends. Connect your carrier over standard SIP or WebSocket, connect your CRM or other systems, then run a pilot list against a success measure you agreed in advance. Getting started with Xebit VoiceCloud walks through each step.
Agree the pass mark first: concurrency, cost ceiling, response time and languages. Then install Xebit VoiceKit on Docker or Kubernetes in your environment, which we do with you the first time. Connect your SIP trunks, your models and your systems, and run it with observability watching every call. Install Xebit VoiceKit covers the requirements and each step.
What do I need before I can make the first call?
Three things: an agent, a SIP trunk or WebSocket connection from a carrier you already use, and an API key if you plan to start calls from your own systems. The number a call comes from is a number on your own trunk. You also need your calling list and script, and, if you want outcomes written back automatically, access to your CRM. See API keys for creating a key.
Can I try Xebit before committing?
Yes. You can hear an agent on your own phone from the Xebit VoiceCloud page before you decide, and we recommend a pilot list with an agreed success measure before you scale. Billing on Xebit VoiceCloud is a monthly minute commitment with a three month minimum term, so talk to us about the pilot before you sign up.
Yes. Tell us the concurrency you need, the cost ceiling you have and the response time you consider acceptable. We set up an environment, you run it on your own hardware against those numbers, and we discuss commercials after your results come back rather than before. See Xebit VoiceKit for what an evaluation involves.
Xebit VoiceCloud vs Xebit VoiceKit
What is the difference between Xebit VoiceCloud and Xebit VoiceKit?
Xebit VoiceCloud is the Xebit platform run by us, priced per connected minute, with compute, the AI models, recordings and storage included. Xebit VoiceKit is the same platform packaged for you to install and run on your own servers, in your own cloud account or fully air gapped, licensed per year by scope rather than by volume. Both include every capability: observability, the audio bridge, conversational intelligence and integrations.
Is Xebit VoiceKit a cut down version of the platform?
No. Xebit VoiceKit is the same artefact we run in Xebit VoiceCloud, not a reduced edition. The licence covers the full stack: Agent SDK, bridge, agent server, speech pipeline, orchestration, scheduling, campaign management, observability, control plane and SIP gateway. Support and all updates are included, with no separate maintenance contract, and it deploys on Docker or Kubernetes into your existing pipeline.
Can I move from Xebit VoiceCloud to Xebit VoiceKit later?
Yes. Because both run the same platform, moving from Xebit VoiceCloud to Xebit VoiceKit is a deployment change rather than a rebuild. Your agents, integrations and the way you call the API carry over; the base URL and where keys are issued change. Teams often start on Xebit VoiceCloud and move once volume passes about 10,000 connected minutes a month or a data requirement appears.
What does running Xebit VoiceKit ask of my team?
You provide the compute, the telephony carrier and, on a standard on prem deployment, your own model licences. In an air gapped deployment the models run locally, so you provide the GPUs instead of the licences. Realistically you also need an engineer who owns the platform. We handle the first deployment and stay on hand while your team takes it over.
Nothing on the infrastructure side while you are on Xebit VoiceCloud, because we run the platform. If you later move to Xebit VoiceKit, you provide compute, your carrier, model licences or, for air gapped, GPUs, and an engineer who owns the deployment. If nobody on your team can own it, Xebit VoiceCloud is usually the better answer and we will say so.
Who applies updates and new releases?
Xebit does. Xebit VoiceCloud is updated for you, so new releases reach your account without any work on your side and your agents, keys and integrations keep working across them. There is nothing to install or upgrade on your side. The release notes list what changed in each release.
Your team does. Each release reaches you as a versioned update, and you apply it on your own infrastructure on your own schedule, after testing it in your environment. All updates and support are included in the Xebit VoiceKit licence. The release notes list what changed in each release.
Calls and agents
How does a Xebit agent handle a call?
Your carrier routes the call over SIP or WebSocket to the Xebit SIP gateway. The bridge routes the audio, the agent server runs the session and the speech pipeline handles the conversation, speech to speech by default. Orchestration enforces turn handling, barge in and your business rules in the call flow. When the call ends, the recording, transcript and outcome are captured and sent to your systems.
Can agents handle both inbound and outbound calls?
Yes. Agents answer inbound calls routed from your carrier and place outbound calls, either one at a time through the API or at volume through campaigns. Scheduling and campaigns handle dialling plans, retries, list management and pacing, so an outbound list runs without someone placing each call. See Creating calls to start a single outbound call with one request.
How quickly does the agent respond?
Under 500 milliseconds on the default speech to speech path, which is fast enough that the conversation feels normal. On the media path alone we measure 118 milliseconds at P95 for a two party call. If you choose a cascade pipeline with your own speech to text and text to speech components, response time depends on the components you bring. Observability records response time on every turn.
Which languages and models does Xebit support?
Speech to speech runs on the Gemini ecosystem by default, and an agent speaks every language that Gemini and OpenAI support in realtime. A cascade pipeline is available when you want to pick each component yourself. On Xebit VoiceKit you bring your own model licences, and in an air gapped deployment you can run local models on GPUs you provide. Language coverage tracks what those providers support.
Can a person take over a call from the agent?
Yes. Live transfer to a human agent is shipped on both Xebit VoiceCloud and Xebit VoiceKit, and the context of the call goes with it. Through the audio bridge, someone on your team can also open a live call in their browser and whisper to one side, join as a third party, or take over completely. The customer stays on the same call throughout, with one recording. See Audio bridge.
What do I get from each call?
Every call leaves a recording, a turn by turn transcript, a short summary, an outcome and the structured fields your workflow needs, such as a promise to pay date or a candidate's notice period. You define the fields with the agent, they are captured during the conversation, and they are sent to your systems as the call ends. Transcripts work in every language the realtime models support.
Telephony and numbers
Does Xebit provide phone numbers or telephony?
No, and that is deliberate. Xebit never provides telephony, phone numbers, SMS, WhatsApp or DLT registration. You keep your own carrier, your own trunks and your own call rates, and Xebit connects to them over standard SIP and WebSocket. Your numbers, routes and contracted rates stay yours, and switching carrier later does not touch your voice application.
How do I connect my carrier?
Point a SIP trunk from your carrier or session border controller at the Xebit SIP gateway, or connect over WebSocket. RTP over SIP, WebRTC, WebSocket and gRPC are all in production, and the bridge converts codecs such as G.711 and Opus automatically. Any carrier that supports standard SIP works. Deployment options explains where the gateway sits in each deployment.
Which number will my outbound calls come from?
A number on your own trunk. When you start a call, the "from" number you pass must be one your carrier has assigned to you and allows you to present. Xebit does not issue or rent numbers, so number registration, caller ID rules and any regulatory filings sit with your carrier. See Creating calls for how the "from" number is set on a request.
Can Xebit send an SMS or WhatsApp message after a call?
Not itself. Xebit does not send SMS, WhatsApp or email to your customers. What it does is fire a webhook when a call ends, carrying the outcome and the fields the agent captured, so your own messaging provider or workflow tool can send the follow up. See Outbound webhooks for the events and payloads you can build on.
Integrations and API
How do I authenticate with the API?
Create an API key in the Xebit VoiceCloud dashboard and send it with each request to https://api.voice.xebit.ai/v1. Keep keys on your server, never in a browser or mobile app, and use separate keys for separate systems so you can revoke one without affecting the others. Auth setup and API keys cover the details.
Create an API key in the Xebit VoiceKit admin console and send it with each request to https://<your-xebit-voicekit-host>/v1, the host your team deployed Xebit VoiceKit on. Keep keys on your server, never in a browser or mobile app, and use separate keys for separate systems. Auth setup and API keys cover the details.
How do I start a call from my own system?
Send one POST request to /v1/calls with the agent, the number to call, the "from" number on your trunk and any variables the agent should use, such as the customer's name or due date. The response identifies the call so you can track it. Any tool that can make an HTTP request can start a call this way. Creating calls has the full request and code samples.
How do I find out when a call has finished?
Subscribe to webhooks. They fire on call lifecycle events, outcomes and transfers, so your endpoint hears about a call as it ends, with the outcome and captured fields. If you prefer to pull, query the call by its ID for its current status and results. See Outbound webhooks and Call status.
Which CRMs and systems does Xebit integrate with?
Zoho and Salesforce connect through the integrations connector, so the outcome, captured fields, summary and transcript land on the right record. Ticketing, loan management, ATS, calendars, data warehouses and workflow tools connect through the REST API and webhooks, and because the orchestration layer is ours, we build the hook into your system as part of deployment. See Workflows and integrations.
How many API requests can I make?
Xebit VoiceCloud applies rate limits to API requests so that one integration cannot crowd out the rest. Build your client to back off and retry when a request is rejected for exceeding the limit, rather than retrying immediately in a loop. Rate limits sets out the limits and how to handle a rejected request.
On Xebit VoiceKit, rate limits are set by your own deployment and licence band, not by Xebit, and your team can tune them to the capacity you run. Build your client to back off and retry when a request is rejected for exceeding the limit. Rate limits explains how limits work and where to configure them.
Data, security and compliance
Does audio leave my network?
On Xebit VoiceCloud, yes. The platform runs in Xebit's infrastructure, so call audio reaches our infrastructure and the model provider that powers the agent, and recordings and transcripts are stored with us. If audio and transcripts must stay inside your network, Xebit VoiceCloud cannot meet that requirement and Xebit VoiceKit is the right choice. See Security for the full data boundary.
It depends on the configuration. On a standard on prem deployment the platform runs inside your network and only inference traffic goes out, to the model provider you hold the licence with. In an air gapped deployment the models run locally on your GPUs and nothing leaves the network. Telephony always runs through your own carrier. See Security.
Is Xebit SOC 2 or ISO 27001 certified?
Not yet. SOC 2, ISO 27001 and an independent VAPT are all in progress, and none is certified today. We would rather say that plainly than show a badge we are not entitled to. If you need the current stage of each for a security review, ask us through the security contact form and we will share where each one stands.
Where are recordings and transcripts stored?
On Xebit VoiceCloud, recordings, transcripts and outcomes are stored in Xebit's infrastructure, and audio storage is included in the per minute rate. You can pull them through the REST API or receive outcomes by webhook whenever your systems need them. Our data retention policy sets out how long each is kept.
On Xebit VoiceKit, recordings, transcripts and outcomes are stored inside your own environment, on storage you provide and control, because the platform runs in your network. Retention is set by your own policies. In an air gapped deployment they never leave the network at all, and your systems read them through the REST API inside that network.
Which security and legal documents can I review?
Our data processing addendum, sub processor list, data retention policy, acceptable use policy, privacy policy, terms of service and business associate agreement are all published and linked from the Security page. For anything beyond those, such as an architecture review or a security questionnaire, use the security contact form to request documentation.
Can the SRE agent change my production setup on its own?
No. The SRE agent detects a problem, opens the incident, correlates signals across layers, writes the root cause analysis and proposes a fix. Nothing changes in production until one of your engineers approves it. If they reject it, the incident stays open with everything the agent found. This holds on Xebit VoiceCloud and Xebit VoiceKit alike. See Observability.
Billing and limits
How much does Xebit VoiceCloud cost?
About $0.052 per connected minute, with a commitment of 5,000 minutes a month, which comes to roughly $258, for a minimum term of three months. One agent is included; if you need more, we scope it with you. The rate covers compute, the AI models, speech to speech and cascade, orchestration, campaigns, observability, recordings, audio storage and human handoff. See Pricing.
How much does Xebit VoiceKit cost?
Xebit VoiceKit is licensed per year by scope, and call volume does not change the price. The SMB band runs from $4,330 to $6,186, Upper SMB from $6,186 to $10,309, and Enterprise from $10,309 upwards, plus a one time setup from $2,062. The licence is paid annually in advance, renewal is flat, and support and all updates are included. See Pricing.
What happens if I use more than my committed minutes?
You top up in blocks of 5,000 minutes at the same $0.052 rate, so going over costs nothing extra per minute. Unused minutes do not carry into the next month, so size the commitment to your expected volume. Above roughly 10,000 connected minutes a month, a Xebit VoiceKit licence usually costs less than renting the minutes.
Nothing changes. Xebit VoiceKit is not metered by minutes, so one call or one million costs the same licence. What limits volume is the capacity of the infrastructure you run it on and the scope of your licence band. In our benchmark a single node held 9,200 concurrent one to one sessions, and in production we sustained 150,000 calls a day on one VM.
Are carrier call charges included?
No. Telephony is the one thing you buy separately, from the carrier you already use, on both Xebit VoiceCloud and Xebit VoiceKit. Your carrier bills you for the calls at the rates you have negotiated with them. The Xebit VoiceCloud per minute rate covers the platform, the AI and storage, and the Xebit VoiceKit licence covers the platform, but neither includes call charges.
How many calls can run at the same time?
Your Xebit VoiceCloud account has a concurrency limit, the number of calls it can run at once, and usage quotas. Plan campaign pacing around the concurrency limit rather than starting every call at the same moment. Concurrency and Quotas explain how each limit applies to your account and what to do when you need more.
Concurrency and quotas are set by your own deployment and licence band, not by Xebit. The platform itself scales with the nodes you run: our published benchmark held 9,200 concurrent one to one sessions per node. Concurrency and Quotas explain how to configure them for your deployment.
Troubleshooting
My calls are not connecting. What should I check?
Start with the call's status and event log, which show where the call stopped. Most connection failures sit between Xebit and the carrier: check that your SIP trunk is pointed at the right gateway address, that credentials or IP allow lists match, and that the "from" number is one your carrier lets you present. Then confirm with your carrier that the trunk is active. See Call status.
The agent is slow or silent on calls. Why?
When servers look healthy but callers hear silence, the cause is usually a model or speech provider timing out or slowing down. Observability tracks response time per turn, dead air and errors for each LLM, speech to text and text to speech provider, and the SRE agent opens an incident naming which provider moved first. Check the open incidents and replay an affected call. See Observability.
My webhooks are not arriving. What should I check?
Check that your endpoint is reachable over HTTPS from where the platform runs, returns a success status quickly and does not block the request with authentication you have not configured. Slow endpoints should acknowledge first and process afterwards. On Xebit VoiceKit, also check that your network allows traffic from the platform to the endpoint. See Outbound webhooks.
My API key has stopped working. What do I do?
First check that the key is being sent to the right base URL: a Xebit VoiceCloud key works only against https://api.voice.xebit.ai/v1, and a Xebit VoiceKit key only against your own host. Then check whether the key was revoked or rotated in the Xebit VoiceCloud dashboardXebit VoiceKit admin console. If it was, create a new key and update every system that used the old one. See API keys.
The agent is not doing what I configured. What should I check?
Replay a recent call with its transcript and event log side by side to see where the conversation went off course. Check whether a recent change to the agent introduced the problem, since observability flags prompt regressions, off script answers and failed tool calls. Then test the fix on a few calls before running the full list again. If a tool call failed, check the system it calls.