AEGIS: An Open-Source, Self-Hosted Personal AI Orchestration System

A developer has open-sourced AEGIS, an MIT-licensed personal AI orchestration platform built and refined over a year for a single user. Rather than a chatbot, it runs as a fleet of scheduled and event-driven workflows split across four permission-scoped agents: Sebas for GTD tasks, Raphael for research, Maou for finances, and Pandora's Actor for infrastructure. The system watches routine signals — tasks, email, money, homelab alerts — and only surfaces a decision when a human genuinely needs to make one.

The stack pairs FastAPI, Postgres with pgvector, and Temporal, deployed on a small Docker Swarm cluster. Model calls go through a LiteLLM proxy exposing fast, balanced, and smart tiers, defaulting to local models and escalating to Claude or GPT only when a task demands it.

The standout design choice is collapsing every human-in-the-loop moment into one primitive: a Postgres row, a chat-app card, and a Temporal workflow that durably waits — sometimes for days — for a response, surviving restarts and redeploys. Agent capabilities and routing live in database metadata rather than hardcoded logic, making the whole system forkable without touching Python. The project also explicitly handles a subtle reasoning-model failure mode where hidden reasoning tokens can silently truncate output.

AEGIS isn't a hosted SaaS or a generic agent framework — it's a complete, opinionated application meant to be forked and configured for someone else's life. For engineers, the interesting part is the pattern: treating every interruption as a durable, auditable event, and modeling agent behavior as data instead of code.