Leon 是一个开源的个人 AI 助手项目，从早期的意图识别助手演进为围绕工具、上下文、记忆和代理执行构建的系统。它能理解目标并选择处理方式：智能模式自动决定，受控模式按确定性的原生技能执行，代理模式则可分步规划。Leon 通过真实工具完成搜索、办公、系统操作、媒体处理和编码辅助等任务，并支持逐步加载的计算机使用能力，可操作桌面与浏览器界面并校验结果。它维护分层记忆以记住长期偏好和近期上下文，既能调用本地模型也可接入远程 AI 服务，在隐私与能力之间取得平衡。架构上分为服务端、Web 应用、技能、桥接与 Python 服务，方便按需扩展。
Leon is your open-source personal AI assistant built around tools, context, memory, and agentic execution.
Leon is designed to stay practical, privacy-aware, and grounded in your real environment. It can operate locally, use dedicated tools instead of relying on free-form guessing, and complete tasks from start to finish across deterministic workflows and agent-style execution.
🧠 What Leon Is Today
Leon is no longer just a classic intent-classification assistant like it was for its first release in 2019.
Today, Leon is being built as a more capable assistant that can understand a goal, choose how to handle it, use tools, remember useful information, and recover when something goes wrong.
Leon can run in different ways depending on the task: smart mode chooses for you, controlled mode follows deterministic native skills and actions, and agent mode can plan step by step.
Leon supports native skills for controlled actions and agent skills for SKILL.md-backed workflows.
Leon can use real tools to get work done instead of only replying with plain text.
Leon includes advanced, progressively loaded computer use that can operate desktop and browser interfaces, verify visual outcomes, and run locally or through Satellite without exposing its full tool surface to unrelated turns.
Leon can use context about your environment so answers stay grounded in what is actually happening on your machine and setup.
Leon keeps layered memory so it can remember durable preferences, day-to-day context, and recent discussion context.
Leon supports both local and remote AI providers, which helps balance privacy, control, and capability.
Under the hood, Leon-native skills follow Skills -> Actions -> Tools -> Functions (-> Binaries).
Leon also keeps a compact self-model and a bounded proactive pulse system so it can stay more consistent over time without flooding itself with unnecessary context.
Why?
Privacy matters: Leon can work with local models and local context instead of forcing everything through third-party services.
Grounded behavior matters: Leon prefers explicit tools, context, and memory over vague model-only responses.
Extensibility matters: skills, toolkits, bridges, and binaries make it possible to keep Leon modular.
Open source matters: anyone can inspect the architecture, build on top of it, and help shape where it goes next.
# Clone the repository
git clone https://github.com/leon-ai/leon.git
# Go to the project root
cd leon
# Install pnpm
npm install --global pnpm@latest
# Install dependencies
pnpm install
Run Leon
# Run Leon
pnpm start
Check Your Setup
# Check the setup went well
pnpm run check
By default, Leon runs locally and the app is available on http://localhost:5366.
🏗️ Architecture Snapshot
At a high level, Leon currently consists of:
server/: the main runtime, routing, memory, context management, HTTP API, and agent/controlled execution
app/: the web application
aurora/: UI components and preview environment
skills/: built-in capabilities, split between native/ skills and agent/ skills
bridges/: Node.js and Python bridges plus toolkit definitions and tool runtimes
tcp_server/: Python services used by parts of the runtime stack
core/context/: generated identity and architecture context documents that describe Leon's current behavior
This repository already includes skills and toolkits for areas such as search, productivity, system utilities, media workflows, coding assistance, memory-backed interactions, and voice/audio features.
📚 Documentation Status
The new docs for Leon 2.0 are not ready yet.
For now:
treat this repository as the source of truth for the 2.0 Developer Preview
Leon has been evolving for a long time, but the current 2.0 work is a major transition period.
For a long time, Leon was a smaller assistant project with a simpler architecture. Today, the core is being rebuilt into a much more capable system around tools, memory, context, and agent-style execution. That means a lot of things are still moving, and it makes contribution harder than it will be once the new docs and architecture settle down.
Another important reason is simply time: Leon is still developed largely during spare time. So progress can be uneven, and opening the project more broadly has to be balanced with keeping the direction coherent while the 2.0 Developer Preview is still taking shape.
📖 The Story Behind Leon
Leon started in 2017 and has been active since 2019. If you want the longer backstory, read the story behind Leon.
🔔 Stay Tuned
X / Twitter is the main place where I share Leon progress updates
