At a Glance:
Leon is your open-source personal AI assistant designed around tools, context, memory, and agentic execution, capable of running locally with both smart mode and step-by-step agent planning.
Overview:
Leon is an open-source personal AI assistant that moves beyond single-turn intent classification by combining tools, layered memory, environmental context, and agentic workflows. It can operate in smart mode to choose its own execution path, controlled mode for deterministic native skills, or agent mode for planned step-by-step tasks. The project is designed to run locally with support for local and remote AI providers. The current development focuses on a 2.0 Developer Preview that introduces skills split between native actions and SKILL.md-backed agent workflows, along with bridges, toolkits, and a dedicated TCP server for Python services.
Key Decision Points:
Execution modes balance control and autonomy: Leon offers a "smart" mode for automatic handling, a "controlled" mode for deterministic native skills, and an "agent" mode that plans tasks step-by-step so developers can choose how much agency to give the assistant.
Local-first architecture with provider flexibility: The assistant can work with local models and local context to keep data on-device while still allowing connections to remote AI providers when needed.
Agent skills are driven by SKILL.md: Rather than only hard-coded intent routing, agent skills use SKILL.md files to define workflows, making skill behavior more modular and inspectable.
2.0 Developer Preview is the active branch: The legacy stable version lives on
master, while the current work on tools, memory, context management, and agent execution is on thedevelopbranch with documentation that is not yet finalized.
Core Features:
Multi-mode execution: Operates in smart, controlled, and agent modes to handle tasks either automatically, deterministically, or through planned step-by-step reasoning.
Native and agent skills: Supports native skills for controlled actions and agent skills backed by SKILL.md definitions, structured under Skills → Actions → Tools → Functions (→ Binaries).
Layered memory system: Retains durable preferences, day-to-day context, and recent discussion history so interactions remain grounded and consistent over time.
Environmental context awareness: Uses information about your machine and local setup to keep answers relevant to what is actually happening in your environment.
Local and remote AI provider support: Can connect to local models for privacy-sensitive tasks or remote providers when more capable models are needed.
Bridges and toolkits: Includes Node.js and Python bridges plus toolkit definitions that extend Leon's ability to use real tools across areas such as search, productivity, system utilities, media, coding, memory, and voice.
Use Cases:
Developers and tinkerers who want a personal assistant that can run locally, use explicit tools, and remember context without routing everything through cloud services.
Users who need an assistant that can switch between deterministic controlled actions and flexible agent-planned workflows depending on the task.
Contributors interested in an evolving open-source assistant architecture that combines skills, toolkits, bridges, and an HTTP API in a single codebase.
Open-Source Alternative Value:
Leon's value as an open-source project comes from its explicit focus on local execution, tool-based grounding, and a transparent skill architecture rather than black-box model responses. The assistant can work entirely with local models and local context, and its behavior is defined through inspectable native skills and SKILL.md-backed agent skills. With native skills, agent skills, bridges, toolkits, and a dedicated Python service layer, the project offers contributors a modular codebase where execution modes, memory layers, and environment context are all part of the open design rather than hidden behind a hosted service.
