Architecture

A classic, deliberately simple web app: one Django server, one database, no extra services such as Redis or Celery.

The building blocks

• Browser: the interface consists of Django templates and a little vanilla JavaScript. All files are served by your own server, without a CDN and without a Node build step.
• nginx: part of the Debian package and runs on the same machine. It provides the host name in the home network and TLS, redirects HTTP to HTTPS, serves the static files directly and passes streamed answers through unbuffered. HTTPS is required for the voice features, because browsers only allow the microphone over HTTPS. In Docker and during development the app serves its static files itself (WhiteNoise).
• gunicorn and Django: gunicorn listens on localhost (127.0.0.1) only, with the gthread worker class (default: 2 processes × 8 threads, 300-second timeout), so long streamed answers do not block other requests. Answers are streamed via server-sent events.
• PostgreSQL with pgvector: the only database – for accounts, chats and the vectors of your own documents. No SQLite.
• Worker: a second process from the same code base that indexes documents in the background. It takes its jobs from a table in PostgreSQL and starts further processes to handle several documents at once.
• Provider adapters: one common interface with an adapter for OpenAI-compatible APIs (OpenAI, OpenRouter, LM Studio), one for Anthropic and one for Google Gemini. Implemented with httpx directly against the HTTP APIs.
• MCP client: connects to MCP servers via the official Python SDK and passes their tools on to the models. The client, the connection test and the tool loop in the chat are done.
• MCP server (node): at /mcp/, MultiGPT is itself an MCP server for other programs, with API keys per account. It runs in the same gunicorn processes, without an extra service.

Network and security

• The app runs only in the home network and cannot be reached from the internet. Outgoing traffic is limited to HTTPS requests to the chosen AI providers and, inside the home network, HTTP to LM Studio.
• All settings come from the environment or from /etc/multi-gpt/.env, never from the repository.
• API keys and MCP server credentials are stored encrypted with Fernet in the database and shown in the admin area only by their last four characters.
• Keys and message contents never appear in logs.
• The service runs as its own system user multi-gpt without access to other shares on the server.

Tech