OpenLumara: A Modular, Local-First AI Agent vs OpenClaw & Hermes

Key Takeaways

• Local-first, module-toggleable design. OpenLumara is a from-scratch Python framework for local models. It ships a ~4,000-token prompt and turns memory, shell, and web access into modules you flip on or off — disabled modules load zero Python, so they add zero tokens and zero attack surface.
• Security posture. It can't see your API keys. Arbitrary shell is disabled by default (behind a Docker sandbox module), and web access routes through a controlled tool rather than raw network calls.
• One-command install. Clone the repo and run run.sh (a Windows batch script is included). It builds a virtualenv, then auto-starts a CLI channel and a web UI at localhost:3000, and takes an automatic data backup on first launch.
• Any OpenAI-compatible backend. It targets llama.cpp and koboldcpp but connects to anything OpenAI-compatible — here Ollama. In Settings → API you set the local URL and any placeholder key, then pick your model (e.g. Qwen) and toggle thinking/reasoning.
• Modules are the differentiator. 16 of 23 modules are enabled by default. Toggling web search or memory is one click; in the demo it recorded a to-do into its own memory via a tool call. User-written extension modules let you add your own capabilities — with the goal of the agent eventually writing its own modules.

Commands & Code Mentioned

git clone <openlumara-repo>   # link in the video description
./run.sh                      # from repo root — builds the virtualenv (run.bat on Windows)
localhost:3000                # web UI opens automatically after install