An open-source assistant that uses tools, not just talk

What it does Leon is a personal AI assistant that runs locally and handles tasks through explicit tools, context, and memory rather than free-form text generation. It operates in three modes—smart, controlled, and agent—ranging from deterministic native skills to step-by-step planning workflows. The assistant can interact with your actual environment, remember preferences across sessions, and recover from errors while routing requests through a hierarchy of skills, actions, tools, and functions.

The interesting bit The project is currently rebuilding itself in public: the 2.0 Developer Preview on the develop branch replaces the legacy 2019 intent-classification engine with a modular, agentic core that includes a compact self-model and a “bounded proactive pulse” to keep context from ballooning. It is a spare-time project, so progress is intentionally uneven and the documentation is still catching up to the code.

Key highlights

• Runs fully offline with local AI providers or mixes in remote ones for heavier tasks.
• Architecture splits execution between native/ skills (deterministic) and agent/ skills (workflow-based via SKILL.md).
• Layered memory tracks durable preferences, day-to-day context, and recent discussion history.
• Includes built-in toolkits for search, productivity, system utilities, media, coding help, and voice/audio.
• Core context documents (LEON.md, ARCHITECTURE.md) are treated as the live source of truth since public docs lag behind.

Caveats

• The 2.0 rewrite is active and unstable; public documentation reflects the old architecture, so you must read the repo’s core/context/ files to understand current behavior.
• Contributor onboarding is deliberately limited right now because the architecture is still shifting underfoot.

Verdict Worth watching if you want a hackable, privacy-first assistant that executes rather than chatters; skip it if you need polished docs and a stable release today.