# GenericAgent

> A minimal, self-evolving autonomous agent framework that gives any LLM system-level control of a local computer through 9 atomic tools.

GenericAgent is an open-source, MIT-licensed autonomous agent framework published by lsdefine. According to the README, its core is about 3K lines of code built around 9 atomic tools and a roughly 100-line agent loop. It lets an LLM control a local computer, including the browser, terminal, filesystem, keyboard and mouse, screen vision and mobile devices over ADB.

## What It Is

GenericAgent is a Python agent framework that connects to LLMs such as Claude, Gemini, Kimi and MiniMax. Its design philosophy is to not preload skills but let them evolve. When the agent solves a new task, it records the execution path as a reusable Skill in its memory, so the same task can be recalled directly next time.

The agent can be run through a terminal UI, a Streamlit web UI, a desktop app, or IM bot frontends for Telegram, Discord and Lark/Feishu. The Chinese section also lists WeChat, QQ, WeCom and DingTalk.

## How the Agent Works

The core loop perceives environment state, reasons about the task, executes tools, writes experience to memory, and repeats. The nine tools are code_run, file_read, file_write, file_patch, web_scan, web_execute_js, ask_user, update_working_checkpoint and start_long_term_update. Through code_run, the agent can install Python packages, write scripts, call external APIs or control hardware at runtime.

Memory is layered: meta rules, an insight index, global facts, task skills/SOPs, and a session archive for long-horizon recall.

## Browser Control and Advanced Capabilities

Web tools run through TMWebdriver, a local WebSocket server plus a Chrome extension, which drives a real persistent browser session and preserves logins and cookies. The README reports passing several bot-detection test pages with this approach. These are project-published results.

Advanced capabilities such as web automation, OCR, vision and computer use are unlocked by instructing the agent to set them up, using its pre-installed SOPs, rather than by manual installation.

## Recent Updates

The roadmap lists TUI v3 (2026-05-23), Morphling mode for project-level skill absorption, Goal Hive multi-worker mode, a desktop GUI, Conductor sub-agent orchestration, and Goal mode for time-budgeted loops. A technical report is on arXiv. The latest release listed is desktop-portable-v0.2.0.

## Setup Path

The README recommends Python 3.11 or 3.12 and not 3.14. Install by cloning the repository, creating a uv virtual environment, installing the package with the ui extra, and copying a key template to fill in an LLM API key. A one-line installer is also provided for Windows, Linux and macOS.

## Features
- Self-evolving skill crystallization from completed tasks
- About 3K-line core with a roughly 100-line agent loop
- 9 atomic tools for code, files, web and memory
- Real-browser control via TMWebdriver preserving login sessions
- Layered memory (L0-L4) with SOPs and session archive
- Computer control via keyboard/mouse, screen vision and ADB for mobile
- Terminal UI, Streamlit UI and desktop app frontends
- IM bot frontends for Telegram, Discord and Lark/Feishu
- Goal mode, Goal Hive and Conductor sub-agent orchestration
- Multiple LLM backends including Claude, Gemini, Kimi and MiniMax

## Integrations
Claude, Gemini, Kimi, MiniMax, Telegram, Discord, Lark/Feishu, WeChat, QQ, WeCom, DingTalk, Chrome, Streamlit, ADB

## Platforms
WINDOWS, MACOS, LINUX, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
desktop-portable-v0.2.0

## Links
- Website: https://github.com/lsdefine/GenericAgent
- Documentation: https://github.com/lsdefine/GenericAgent/blob/main/docs/installation.md
- Repository: https://github.com/lsdefine/GenericAgent
- EveryDev.ai: https://www.everydev.ai/tools/genericagent