Skip to content
Star
Open source · MIT · Self-hosted

The open-source voice AI agent that picks up the phone.

Voxos is a self-hosted AI agent you can talk to, call, text and WhatsApp. Teach it skills, put its chores on a schedule and plug in any MCP tool. Your keys, your database, your agent.

Get started Star on GitHub

Python, FastAPI, Postgres and LiveKit. One make up to run it all locally.

VoxosChat with your agent online

  1. Can you call the dentist and move my cleaning to Friday morning?

  2. On it. Calling Dentist now.

  3. DentistOutbound call · saved contact

    “Hi! I'm calling for Sam to move Tuesday's cleaning to Friday morning.”

    “Friday at 10:30 works.”

  4. Nice. Remind me Thursday evening?

  5. Done. I'll ping you here on Thursday at 6 pm.

    Scheduled · Thu 6:00 pm
An example Voxos chat: the agent phones a saved contact to move an appointment, then schedules a reminder.
Voice first

Talk, call, text or WhatsApp. Same agent.

One agent, one memory, one set of tools, whichever way you reach it. Start a thought on a call and finish it in a text.

  • Just talk to it

    Realtime speech-to-speech over LiveKit, using OpenAI, Gemini or Grok voice models. Talk the way you would to a person.

    • Web console or the iOS and Android app
    • Voice and model picked per agent
    • Transcripts saved with your chat history
    Voice docs

  • Give it a phone number

    Connect a Twilio or Telnyx number. Your agent picks up when people call, and it can ring your saved contacts to get something done.

    • Inbound and outbound calls
    • Caller access set per contact
    • Outbound calls with a stated goal
    Phone docs

  • Text it like a friend

    Two-way SMS on the same number. Text a question from anywhere and the reply arrives when the work is done.

    • Replies run as durable background tasks
    • Recent message and call activity
    • Twilio and Telnyx numbers
    SMS docs

  • Drop it into WhatsApp

    Link a WhatsApp Business number as a companion device. Your agent reads direct and group chats and replies right there.

    • Direct and group text messages
    • Approved messages to saved contacts
    • An encrypted session for each user
    WhatsApp docs
Not just chat

It actually does stuff.

Skills, schedules, tools and a memory, so your agent can carry real work from one conversation to the next.

  • Skills

    Teach it a routine once

    Skills are reusable playbooks your agent finds and follows. Edits it suggests wait for your approval.

    Learn more about skills
  • Scheduled tasks

    Put chores on a schedule

    One-off or repeating schedules that post results back to their chat. Ask for a reminder in chat and it sets one up.

    Learn more about scheduled tasks
  • MCP tools

    Plug in any MCP server

    Remote MCP over Streamable HTTP or SSE, with OAuth, bearer or API-key auth. You pick exactly which tools each agent gets.

    Learn more about mcp tools
  • Browser tasks

    Let it drive a browser

    A built-in local browser, Browserless or Browser Use, with website logins kept encrypted in your vault.

    Learn more about browser tasks
  • Memory

    It remembers stuff

    Durable memories, persona rules and task plans, plus suggested improvements you review before they stick.

    Learn more about memory
  • Subagents

    It delegates the busywork

    Agents can hand work to subagents, and a subagent never gets more tool access than its parent.

    Learn more about subagents
  • Webhooks

    Start work from anywhere

    Agent webhooks start a durable run from any service that can send an authenticated HTTP request.

    Learn more about webhooks
  • Local models

    Bring your own model

    Point text chat at any OpenAI-compatible endpoint, including Ollama, LM Studio, vLLM and llama.cpp.

    Learn more about local models

Example skills to steal

11 curated skills, each a bounded playbook built around real MCP tools. Load them into your account and edit them to taste.

Browse the skills
Apps included

A web console and a phone app.

Set everything up in the web console, then chat or call your agent from the iOS and Android app. Both connect to your own Voxos server.

The web console: chat, voice, phone numbers, MCP servers, schedules and more.
The Flutter app for iPhone and Android, with one-tap voice calls.
Integrations

Plays nicely with the stuff you already use.

Bring your own models, phone numbers and tools. Every chip links to its setup guide.

Models

Realtime voice and text models, hosted or on your own hardware.

Calls and messaging

Real phone numbers, realtime media and the chat app everyone already has.

MCP catalog

Tools your agent can pick up, each with setup notes.

Runs on

Boring, dependable infrastructure you control.

Anything else that speaks MCP over Streamable HTTP or SSE works too. Add your own server.

How it works

Your agent, on your infrastructure.

Voxos is the backend in the middle. It routes every channel to the same agent, keeps state in your Postgres, and calls out only to the providers you configure.

  1. You

    • Voice in the web console or app
    • Phone calls
    • SMS
    • WhatsApp
    • Text chat
    • Webhooks and schedules

  2. Voxos

    • FastAPI API
    • LiveKit voice workers
    • Durable run worker and scheduler
    • WhatsApp worker
    • Postgres + pgvector, row-level security
    • Encrypted secret vault

  3. Your providers

    • OpenAI, Gemini, xAI or local models
    • Twilio or Telnyx numbers
    • Remote MCP servers
    • Browser automation
Get started

Up and running in a coffee break.

Clone the repo, add a few secrets to .env, and bring up the whole stack with Docker Compose.

~/code 
git clone https://github.com/s-block/voxos.gitcd voxoscp .env.example .env# add the secrets from the Quick Start to .env
make up# API, LiveKit, Postgres + pgvector and the workers# then open the console at http://localhost:5173
FAQ

Questions people actually ask.

What is Voxos?

Voxos is an open-source, self-hosted backend for voice-first AI agents. You can talk to your agent in real time, call it or have it call people, text it over SMS, chat with it on WhatsApp, and give it skills, schedules, memory and MCP tools. It is MIT licensed.

Read the docs →

Is Voxos free?

Yes. The code is MIT licensed and you run it on your own machine or cluster. You pay your own providers directly for things like model APIs, phone numbers and hosting.

MIT License →

Can Voxos answer and make phone calls?

Yes. Connect a Twilio or Telnyx number and inbound calls reach your agent through LiveKit. Agents can also call your saved contacts with a stated goal, and caller access is set per contact.

Phone integrations →

Does Voxos work with WhatsApp?

Yes. Voxos links a WhatsApp Business number as a companion device, so your agent can receive direct and group text messages and reply in the chat. Each user can link one number.

WhatsApp setup →

Which AI models does Voxos support?

Realtime voice runs on OpenAI, Google Gemini and xAI Grok voice models. Text chat supports OpenAI, xAI, the NVIDIA API Catalog and any OpenAI-compatible endpoint.

Providers and keys →

Can I use local or self-hosted models?

Yes, for text chat. Point Voxos at any OpenAI-compatible endpoint, such as Ollama, LM Studio, vLLM or llama.cpp, on your machine or your network.

Self-hosted models →

How does Voxos use MCP?

The Model Context Protocol lets an agent use external tools. Voxos connects to remote MCP servers over Streamable HTTP or SSE, handles OAuth and API keys, and only gives an agent the tools you explicitly allow.

MCP and tools →

How does Voxos keep keys and data safe?

Provider keys and tokens are encrypted per user, with AES-GCM for local development and AWS or Google Cloud KMS envelope encryption for deployments. Postgres row-level security keeps each user's data separate, and agents only use the tools they have been granted.

Architecture and security →

What do I need to run Voxos?

Docker with Docker Compose, a Firebase project for sign-in, and API keys for the providers you want to use. Running make up starts the API, LiveKit, Postgres with pgvector and the workers. A Helm chart covers Kubernetes.

Quick Start →

Is there a mobile app?

Yes. A small Flutter app for iPhone and Android gives you chat and one-tap voice calls with your agent. It connects to your own Voxos server.

Mobile app →

Give your agent a voice.

Self-host Voxos, connect a number, and say hello. The whole thing is MIT licensed and yours to change.

Read the Quick Start Star on GitHub