OpenLumara
A modular, token-efficient AI agent framework, written from scratch in Python, built for local models with a small system prompt and hard, code-enforced limits on what the AI can do, rather than relying on prompt engineering for safety.
OpenLumara keeps its system prompt as small as around 4,000 tokens in normal use, which keeps token usage down both with local models served through llama.cpp or koboldcpp and with public model APIs. Every component beyond the minimal core, including shell access, memory, the scheduler, time-awareness and token-awareness, is a module that can be turned on or off with a /module command, down to an empty system prompt that talks to the base model directly.
It connects to any OpenAI API-compatible backend, local or cloud, and can be self-hosted for full privacy or run on a cloud server. It ships interfaces for a browser-based WebUI, a terminal CLI, Telegram, Discord and Matrix (with encryption support), with more planned.
The project's core framework was designed and hand-coded by its author, who used AI only to ask how to fix bugs or accomplish specific things in Python, auditing and modifying anything AI-assisted before it went in; some non-core channel integrations were mostly AI-generated and then manually audited.