OpenLumara: A lightweight, secure AI agent built from scratch for local models
4K-token system prompt, modular design, and no vibecoding – runs fast on modest hardware.
OpenLumara, created by developer Rose22, is a new AI agent framework written entirely from scratch (no vibecoding) and optimized for running local models on modest hardware. Its default system prompt is around 4K tokens—far smaller than typical agents—and every feature is modular and can be disabled independently. Security is built in from the ground up, with all actions controlled via toolcalls that restrict what the AI can see or do, avoiding the shell-access vulnerabilities found in projects like OpenClaw.
The WebUI is a priority and has been hand-crafted for usability. The project is fully open-source under GPL2, with no commercial interests. Rose22 manually coded all core security components and only used AI assistance for the web interface and some parsing logic (like toolcall stream parsing), with every line vetted and commented. OpenLumara is already integrated into esolithe’s KoboldCpp fork (esobold). A video demo shows it managing calendars and todos as a personal assistant. Ideal for LLM enthusiasts who want a lightweight, secure, and fast agent on local hardware.
- Default system prompt is ~4K tokens, extremely token-efficient for local models.
- Modular design: everything from tools to subsystems can be turned on/off independently.
- Security-first: toolcall-based control, no shell access needed, GPL2 open-source license.
Why It Matters
A lean, locally-runnable agent that prioritizes security and performance over bloat, challenging the status quo.