Buyer's guide ยท Vox
AI voice assistant source code you can run as a branded subscription service
VOX is self-hosted source code for a voice assistant that calls its user: a spoken morning briefing, wake-up calls and reminders, research with a spoken summary, important-mail alerts and confirmed PC tasks. The engine is Python 3.11 with a Node.js 20 command line for onboarding. It runs for one person, or in SaaS mode as a subscription service under your own name, with sign-up, subscriber consoles, an operator panel and Stripe billing. Calls go through your own Twilio or Telnyx account.
What you get
The full source code
The Python voice engine (calls, jobs, rituals, channels, console), the Node.js onboarding CLI, an offline test suite that uses fake providers, and 16 pages of documentation.
Calls with barge-in
The assistant places a call through Twilio or Telnyx and holds a two-way conversation. The user can interrupt it.
Daily routines
A spoken morning briefing, wake-up calls, reminders and check-ins with time zones and quiet hours, plus a 12-question onboarding that builds the user's profile.
Research and agent tasks
Ask for research and get a short spoken result with sources. Multi-step tasks use an orchestrator with up to four read-only subagents and hard token and cost limits; actions come back as proposals only.
Mail and calendar
Read-only Gmail triage that can trigger a call for important mail and flags phishing. Google Calendar and MCP tools are read, and anything that writes needs confirmation.
PC actions with undo
Seventeen file and system actions with a plan, a confirmation step, a journal and undo. It never deletes. Mouse and keyboard control is off by default.
Channels to the owner
SMS, WhatsApp, e-mail, text commands and an inbound call protected by a PIN. WhatsApp works through Twilio only.
A console and a setup wizard
A local console with live key tests, calls, jobs, rituals, logs and an audit log, plus data export and delete for GDPR.
Docker and installer
A Dockerfile, docker-compose, an install script, systemd units and Caddy and nginx examples.
How an agency or founder runs it as a service
SaaS mode is part of the source code. The same engine serves many subscribers under your brand.
A public site and sign-up
A public site, subscriber sign-up and a console for each tenant, plus an operator panel for you.
Plans and credit packs in your Stripe
Subscriptions and prepaid credit packs are paid to your own Stripe account. No plan or price ships by default: you create them.
Usage quotas per plan
Limits on call minutes, research jobs, texts, rituals and messages per plan, so you can keep provider costs under plan revenue.
A margin view
The operator side shows provider cost against revenue, with MRR, churn and ARPU worked out from recorded rows. Cost per call is an estimate from a per-minute model, not an invoice.
Your own brand
Your name, colours and legal pages in the branding settings.
Without Stripe or SMTP
You can still assign plans by hand without Stripe. Without an SMTP server, new accounts wait for your activation.
What it needs to run
None of these services is included in the price. Each provider bills you directly.
- A Linux server or VPS with Docker (host networking is used), or Python 3.11+ and Node.js 20+ for the manual install
- A domain or sub-domain with HTTPS through Caddy or nginx; the phone provider connects to it over a secure WebSocket
- Your own Anthropic API key for the language model
- Your own ElevenLabs API key and a voice ID, or your own Piper and whisper.cpp servers
- A Twilio or Telnyx account with a voice-capable number
- Optional: Stripe for billing, an SMTP server, a Google OAuth client for Gmail and Calendar
What it does not do
- No real call, text, Stripe charge, Google sign-in or live model request has been run against it. The 1,900+ automated tests use fake providers and signed fake webhooks. Expect to test it on your own accounts before you sell it.
- Calls to anyone but the verified owner start with an AI disclosure that cannot be switched off. Cold calling is refused in code: there is no bulk dialler and no list import. Calls are limited to 3 per hour and texts to 6 per hour and 30 per day.
- One risk is still open: whether ElevenLabs marks the end of the disclosure audio on a live line has not been verified live. If it does not, third-party calls would hang up after the disclosure.
- Phone numbers, any telecom registration in your country and legal advice on calling rules are not included. You are responsible for the calls your install places.
- Optional mouse and keyboard control and local speech (Piper, whisper.cpp) were tested with fakes only.
- There is no native mobile or desktop app, wake word or push notification.
- Docker Desktop on macOS or Windows is unverified, because the Docker recipe relies on host networking, a Linux feature.
- The live demo runs SaaS mode on demo data with no provider key, so it places no call.
Questions
Can I use it to call my customers or run outbound campaigns?
No. Campaign and cold dialling are not implemented. A call to the user's own verified number goes straight to the assistant. A call to any other number starts with a fixed AI disclosure, and third-party callees get no tools. There is no bulk dialler or list import.
Is it tested on real phone calls?
Not yet. The test suite is offline and uses fake providers. No real call, SMS, Stripe charge or live model request has been made, so test it with your own Twilio or Telnyx and ElevenLabs accounts first.
Which AI and phone providers does it use?
Anthropic for the model, ElevenLabs for speech (or your own Piper and whisper.cpp servers), and Twilio or Telnyx for calls. You use your own accounts and pay them per minute and per token.
Do I need Stripe to run it?
Only if you want to charge subscribers. Without Stripe the operator can assign plans by hand. Stripe billing was tested with signed fake webhooks, never against live Stripe.
Can I use it for myself only?
Yes. The single-user install runs a local console and phones one verified number, your own. SaaS mode is a separate part of the same code.
How much does Vox cost, and what does the licence allow?
$139, one payment, no subscription. One licence, no tiers: any number of projects and clients, a paid SaaS, every line changed. The only limit is reselling or publishing the source code itself. The services you connect bill you directly. See the licence.
Is setup included?
Yes. We install Vox for you and put in your logo, colours and texts. Bigger changes or new features get a fixed price before we start.
What if it does not work?
Write to us within 14 days. We fix it, usually the same day. If we cannot fix it within 7 days, you get your money back. See the refund policy.