A minimal (~3K lines), self-evolving autonomous agent framework that grows its own skill tree from execution history. It provides system-level control over a local computer, using a ~100-line core loop and 9 atomic tools while operating within a <30K token context window.
Project overview
The framework distinguishes itself through a ~3K-line architecture with a ~100-line agent loop that automatically converts successful task executions into permanent reusable skills, reducing token consumption compared to frameworks requiring 200K-1M token context windows.
Project type
AI Agent
Use cases
Automation
Deployment
Refer to project documentation
License
MIT
Best for
Developers and AI engineers looking for a self-hosted agent framework that retains learned skills from execution history while operating within a low-token context window.
Users needing to execute system-level control tasks involving terminal, filesystem, keyboard/mouse input, screen vision, and mobile devices via ADB.
Key capabilities
A ~100-line core loop that enables an LLM to perceive environment state, reason about tasks, execute tools, and write experience to memory.
Automatically crystallizes each successful task execution into a reusable Skill, growing a personal skill tree from seed code without manual intervention.
Provides 9 atomic tools including code_run, file_read, file_write, file_patch, web_scan, web_execute_js, ask_user, and memory update functions for foundational interaction.
Features a 5-layer (L0-L4) hierarchical memory system including Meta Rules, Insight Index, Global Facts, Task Skills, and Session Archive for efficient routing and recall.
Injects into a real browser via a local WebSocket server and Chrome extension to preserve login sessions, cookies, and normal browser fingerprints, passing bot detection tests.
Grants system-level control over a local computer covering terminal, filesystem, keyboard/mouse input, screen vision, and mobile devices (ADB).
Dynamically installs Python packages, writes scripts, calls external APIs, or controls hardware at runtime, crystallizing temporary abilities into permanent tools.
Multi-worker cooperative Goal mode using BBS-coordinated master/workers running long-horizon objectives in parallel.
Limitations and risks
The Terminal UI rendering on Windows can be flaky depending on the terminal and font used.
The framework is incompatible with Python 3.14 due to dependencies such as pywebview.
Pre-installed SOPs are written in Chinese, requiring a manual translation prompt for English-speaking users.
The framework provides direct system-level control, including arbitrary code execution, which can be dangerous if misused.
Getting started
Setup is considered easy due to a one-line installer for Windows, Linux, and macOS, alongside clear clone and install instructions. The first success path involves cloning the repository, installing with uv, adding an API key, and running the framework.