DEV Community

Andres Sarrazola
Andres Sarrazola

Posted on Fully Autonomous

OpenLivery: self-hosting multi-tenant WhatsApp AI agents for agencies

#opensource

I’m the developer of OpenLivery, an MIT-licensed platform for agencies managing WhatsApp AI agents across multiple clients.

The core workflow is one workspace per client, managed from one installation. That matters when each client needs their own knowledge base, conversations and branded portal.

What is included

The repository contains the application, rather than just an agent prompt or workflow template:

  • A Next.js frontend and a FastAPI backend.
  • PostgreSQL, database migrations and Docker Compose deployment.
  • Client workspaces, knowledge bases, branded portals and human handoff.
  • WhatsApp Cloud API, a QR-based WhatsApp bridge and a web chat widget.
  • OpenAI, Anthropic and OpenAI-compatible model providers.
  • Per-agent HTTP tools and MCP servers for extending what agents can do.

Start a local installation

With Git, Docker Compose and make available, the repository’s quick start is:

git clone https://github.com/sarrazola/openlivery.git
cd openlivery
./scripts/generate-docker-env.sh
make up
Enter fullscreen mode Exit fullscreen mode

Follow the prompts from the environment generator and the current README for configuration. The local web interface runs on port 3000 and the API on port 8000.

Starting the containers is the first step. You still need to configure your model provider and connect the messaging channel you intend to use. Self-hosting does not remove infrastructure, model or messaging-provider costs.

Try the agency workflow

A useful evaluation is to create two test client workspaces, give each different knowledge content, and run separate conversations. Then try human takeover and the client portal for each workspace. This makes the client-management workflow easier to assess than a single chatbot demo. Use test data while evaluating the setup.

For tools, start with one narrow operation that the client actually needs. Configure its HTTP integration or MCP server for that agent, then test success and failure responses before adding more capabilities.

Why publish the full stack?

Agencies should be able to inspect and adapt the platform they run for clients. The MIT license allows them to self-host and modify the code, including their branding and integrations. English and Spanish documentation is included in the repository.

This is an early project, and feedback on installation and client onboarding is especially useful. If you run agents for several clients, which part is hardest today: setup, knowledge maintenance, integrations or handing conversations to a person?

Source code and installation instructions · Website

Disclosure: AI tools supported development and helped draft this article.

Top comments (0)