Leon

Leon is an open source personal AI assistant built around tools, context, memory and agentic execution. It began in 2019 as an intent classifier and now aims at goals rather than commands, deciding how to handle a request, calling real tools instead of guessing in prose, and recovering when a step goes wrong.

Work runs in one of three modes. Smart mode picks for you, controlled mode follows deterministic native skills and actions, and agent mode plans step by step. Native skills resolve down through actions, tools and functions, while agent skills are backed by SKILL.md files.

Several design choices keep the assistant tied to the machine it runs on.

• Local or remote models: both provider kinds are supported, so privacy and capability can be balanced per setup.
• Layered memory: durable preferences, day-to-day context and recent discussion are kept apart rather than merged.
• Environment context: answers draw on what is actually happening on your machine instead of assumptions.
• Computer use: progressively loaded desktop and browser control that verifies visual outcomes without exposing its whole tool surface to unrelated turns.
• Extensibility: skills, toolkits, bridges and binaries keep the parts modular and replaceable.

Setup is a clone, a pnpm install and a start command, on Linux, macOS or Windows with Node.js 24 or newer. Development currently centers on a 2.0 developer preview whose documentation is unfinished, so anyone wanting the more stable earlier architecture should use the master branch.