Rose22/openlumara

handwritten harness for AI, not vibecoded, everything is a module/plugin, tailor made for use with local models, tiny system prompt (around 1k by default), support for exclusive llamacpp features, strong security, strict zero-trust-in-the-ai policy (the code hard limits the ai, no amount of prompt engineering will bypass it)

What it solves

OpenLumara is a lightweight, local-first AI agent framework designed to reduce token consumption and overhead. It provides a modular system for managing personal tasks, notes, and routines, making it particularly useful for individuals with executive dysfunction (such as ADHD or autism) who need a structured AI assistant.

How it works

The framework is written in Python and connects to any OpenAI API-compatible backend, including local providers like llamacpp, ollama, and koboldcpp. It uses a modular architecture where features—such as memory, scheduling, and shell access—can be toggled on or off to minimize the system prompt size. It supports multiple communication channels including a WebUI, CLI, Telegram, Discord, and Matrix.

Who it’s for

  • Local AI enthusiasts: Users who want a private, self-hosted agent that is fast and token-efficient.
  • Developers: Those looking for a modular framework that is easy to extend via Python classes for new modules or channels.
  • Personal productivity seekers: People managing life tasks, habits, and routines through an AI interface.

Highlights

  • Extreme Token Efficiency: Optimized system prompts (as low as 4000 tokens) and tools to monitor real-time token usage.
  • Modular Design: Almost every core component (memory, scheduler, time-awareness) can be disabled to reduce overhead.
  • Sandboxed Shell: A toggleable module that runs commands inside a Docker or Podman container for security.
  • Multi-Channel Support: Native integration with WebUI, CLI, Telegram, Discord, and Matrix.
  • Integrated Scheduler: A custom-built system for scheduling AI tasks.
  • Memory System: Uses MessagePack format for compact and fast data storage.

Related

  • Project
    tinyhumansai/openhuman

    OpenHuman is an open‑source Rust‑based AI‑agent harness offering a desktop app and a library. It runs hundreds of agents on a cheap VPS by compressing prompts, using a tiny tool‑search model, and keeping the core in a single compiled process. Benchmarks show it’s the fastest, cheapest, and most memory‑efficient among seven comparable harnesses. Features include built‑in memory, voice agents, real‑browser/desktop control, 119 OAuth integrations, 14 messaging channels, and support for local or BYOK models. Install via pre‑built binaries or a simple script; developers embed it with a few lines of Rust code.

  • Project
    open-jarvis/OpenJarvis

    A local-first framework for building and deploying personal AI agents on personal devices, reducing cloud dependency by optimizing for on-device hardware constraints.

  • Project
    amaiya/onprem

    OnPrem.LLM is a privacy-focused Python toolkit for applying LLMs to sensitive documents, supporting local execution with optional cloud backends, RAG, structured outputs, and sandboxed agents.

  • Project
    luoguoxiong/aipack

    A lightweight, extensible TypeScript Agent framework providing a runtime and pluggable ecosystem for building AI agents with built-in support for memory, context compression, and MCP.