278 lines
10 KiB
Markdown
278 lines
10 KiB
Markdown
# GenericAgent — 3,300 Lines to Full OS Autonomy
|
||
|
||
[English](#english) | [中文](#chinese)
|
||
|
||
<a name="english"></a>
|
||
|
||
A minimalist autonomous agent framework that gives any LLM physical-level control over your PC — browser, terminal, file system, keyboard, mouse, screen vision, and mobile devices — in ~3,300 lines of Python.
|
||
|
||
No Electron. No Docker. No Mac Mini. No 500K-line codebase. No paid installation service.
|
||
|
||
## See It in Action
|
||
|
||
<p align="center">
|
||
<img src="assets/demo/order_tea.gif" alt="Agent ordering milk tea" width="600">
|
||
<br><em>"Order me a milk tea" — the agent navigates a delivery app, picks items, and checks out, fully autonomously.</em>
|
||
</p>
|
||
|
||
<table>
|
||
<tr>
|
||
<td width="50%"><img src="assets/demo/autonomous_explore.png" width="100%"><br><em>Autonomous quantitative analysis — the agent explores data sources and generates reports on its own schedule.</em></td>
|
||
<td width="50%"><img src="assets/demo/wechat_batch.png" width="100%"><br><em>WeChat batch messaging — yes, it can drive WeChat too.</em></td>
|
||
</tr>
|
||
</table>
|
||
|
||
## What Happens When You Use It
|
||
|
||
```
|
||
You: "Read my WeChat messages"
|
||
Agent: installs dependencies → reverse-engineers DB → writes reader script → saves as SOP
|
||
Next time: instant recall, zero setup.
|
||
|
||
You: "Monitor stock prices and alert me"
|
||
Agent: installs mootdx → builds screening workflow → sets up scheduled task → saves as SOP
|
||
Next time: one sentence to run.
|
||
|
||
You: "Send this file via Gmail"
|
||
Agent: configures OAuth → writes send script → saves as SOP
|
||
Next time: just works.
|
||
```
|
||
|
||
**Dogfooding**: This repository — from installing Git to `git init`, writing this README, to every commit message — was built entirely by GenericAgent without the author opening a terminal once.
|
||
|
||
Every task the agent solves becomes a permanent skill. After a few weeks, your instance has a unique skill tree — grown entirely from 3,300 lines of seed code.
|
||
|
||
## The Seed Philosophy
|
||
|
||
Most agent frameworks ship as finished products. GenericAgent ships as a **seed**.
|
||
|
||
The 5 core SOPs define how the agent thinks, remembers, and operates. From there, every new capability is discovered and recorded by the agent itself:
|
||
|
||
1. You ask it to do something new
|
||
2. It figures out how (install dependencies, write scripts, test)
|
||
3. It saves the procedure as a new SOP in its memory
|
||
4. Next time, it recalls and executes directly
|
||
|
||
The agent doesn't just execute — it **learns and remembers**.
|
||
|
||
## Quick Start
|
||
|
||
```bash
|
||
# 1. Clone
|
||
git clone https://github.com/lsdefine/pc-agent-loop.git
|
||
cd pc-agent-loop
|
||
|
||
# 2. Install minimal deps
|
||
pip install streamlit pywebview
|
||
|
||
# 3. Configure API key
|
||
cp mykey_template.py mykey.py
|
||
# Edit mykey.py with your LLM API key
|
||
|
||
# 4. Launch
|
||
python launch.pyw
|
||
```
|
||
|
||
**Also runs on Android** — tested successfully on Termux with `python agentmain.py` (CLI frontend):
|
||
|
||
```bash
|
||
# In Termux
|
||
cd /sdcard/ga
|
||
python agentmain.py
|
||
```
|
||
|
||
Once running, tell the agent: *"Execute web setup SOP to unlock browser tools"* — it handles the rest. See [WELCOME_NEW_USER.md](WELCOME_NEW_USER.md) for the full bootstrap sequence.
|
||
|
||
## vs. Alternatives
|
||
|
||
| | GenericAgent | OpenClaw | Claude Code |
|
||
|---|---|---|---|
|
||
| Codebase | ~3,300 lines | ~530,000 lines | Open-source (large) |
|
||
| Deploy | `pip install` + API key | Multi-service orchestration | CLI + subscription |
|
||
| Browser | Injects into real browser (keeps login state) | Sandboxed/headless | Via MCP plugins |
|
||
| OS Control | Keyboard, mouse, vision, ADB | Multi-agent delegation | File + terminal |
|
||
| Self-evolution | Grows SOPs & tools autonomously | Plugin ecosystem | Stateless per session |
|
||
| Core shipped | 10 .py + 5 SOPs | Hundreds of modules | Rich CLI toolkit |
|
||
|
||
## How It Works
|
||
|
||
```
|
||
User instruction
|
||
↓
|
||
┌─────────────────────┐
|
||
│ agent_loop.py (92L) │ ← Sense-Think-Act cycle
|
||
│ "What do I know? │
|
||
│ What should I do?" │
|
||
└────────┬────────────┘
|
||
↓
|
||
┌─────────────────────┐
|
||
│ 7 Atomic Tools │ ← All capabilities derive from these
|
||
│ code_run │ Execute any Python/PowerShell
|
||
│ file_read/write │ Direct disk access
|
||
│ file_patch │ Surgical code edits
|
||
│ web_scan │ Read live web pages
|
||
│ web_execute_js │ Control browser DOM
|
||
│ ask_user │ Human-in-the-loop
|
||
└────────┬────────────┘
|
||
↓
|
||
┌─────────────────────┐
|
||
│ Memory System │ ← Persistent across sessions
|
||
│ L0: Meta-SOP │ How to manage memory itself
|
||
│ L2: Global Facts │ Environment, credentials, paths
|
||
│ L3: Task SOPs │ Learned procedures (self-growing)
|
||
└─────────────────────┘
|
||
```
|
||
|
||
The agent starts with 7 primitive tools. Through `code_run`, it can install packages, write scripts, and interface with any hardware or API — effectively manufacturing new tools at runtime.
|
||
|
||
<details>
|
||
<summary>What Ships in the Box</summary>
|
||
|
||
**Core engine** (runs the agent):
|
||
- `agent_loop.py` — Sense-Think-Act loop (92 lines)
|
||
- `ga.py` — Tool definitions and execution
|
||
- `sidercall.py` — LLM communication (multi-backend)
|
||
- `agentmain.py` — Session orchestration
|
||
|
||
**Interface** (talk to the agent):
|
||
- `stapp.py` — Streamlit web UI
|
||
- `tgapp.py` — Telegram bot interface
|
||
- `launch.pyw` — One-click launcher with floating window
|
||
|
||
**Infrastructure**:
|
||
- `TMWebDriver.py` — Browser injection bridge (not Selenium — injects JS into your real browser via Tampermonkey)
|
||
- `simphtml.py` — HTML→text cleaner for web perception
|
||
|
||
**5 Core SOPs** (shipped, version-controlled):
|
||
1. `memory_management_sop` — L0 constitution: how the agent manages its own memory
|
||
2. `autonomous_operation_sop` — Self-directed task execution
|
||
3. `scheduled_task_sop` — Cron-like recurring tasks
|
||
4. `web_setup_sop` — Browser environment bootstrap
|
||
5. `ljqCtrl_sop` — Desktop physical control (keyboard, mouse, DPI-aware)
|
||
|
||
Everything else — Gmail integration, WeChat automation, vision APIs, game downloaders, stock analysis workflows — the agent builds and memorizes on its own through use.
|
||
|
||
</details>
|
||
|
||
---
|
||
|
||
<a name="chinese"></a>
|
||
|
||
# GenericAgent — 3,300 行代码，完整 OS 级自主控制
|
||
|
||
一个极简自主 Agent 框架。用约 3,300 行 Python，让任意 LLM 获得对你 PC 的物理级控制能力——浏览器、终端、文件系统、键鼠、屏幕视觉、移动设备。
|
||
|
||
不需要 Electron，不需要 Docker，不需要 Mac Mini，不需要 53 万行代码，不需要付费安装服务。
|
||
|
||
## 用起来是什么样的
|
||
|
||
```
|
||
你："帮我读取微信消息"
|
||
Agent：安装依赖 → 逆向数据库 → 写读取脚本 → 保存为 SOP
|
||
下次：一句话直接调用，零配置。
|
||
|
||
你："帮我监控股票并提醒"
|
||
Agent：安装 mootdx → 构建选股工作流 → 设置定时任务 → 保存为 SOP
|
||
下次：一句话启动。
|
||
|
||
你："用 Gmail 发这个文件"
|
||
Agent：配置 OAuth → 写发送脚本 → 保存为 SOP
|
||
下次：直接能用。
|
||
```
|
||
|
||
**自举实证**：本仓库从安装 Git、`git init`、编写 README 到每一条 commit message，全程由 GenericAgent 完成——作者没有打开过一次终端。
|
||
|
||
每个解决过的任务都会变成永久技能。用几周后，你的 Agent 实例会拥有一套独特的技能树——全部从 3,300 行种子代码中生长出来。
|
||
|
||
## 自举哲学
|
||
|
||
多数 Agent 框架以成品形态发布。GenericAgent 以**种子**形态发布。
|
||
|
||
5 个核心 SOP 定义了 Agent 如何思考、记忆和行动。之后的一切能力，由 Agent 在使用中自主发现并记录：
|
||
|
||
1. 你让它做一件新事
|
||
2. 它自己摸索方法（安装依赖、写脚本、测试）
|
||
3. 把流程保存为新 SOP
|
||
4. 下次直接调用
|
||
|
||
Agent 不只是执行——它**学习并记忆**。
|
||
|
||
## 快速开始
|
||
|
||
```bash
|
||
# 1. 克隆
|
||
git clone https://github.com/lsdefine/pc-agent-loop.git
|
||
cd pc-agent-loop
|
||
|
||
# 2. 安装最小依赖
|
||
pip install streamlit pywebview
|
||
|
||
# 3. 配置 API Key
|
||
cp mykey_template.py mykey.py
|
||
# 编辑 mykey.py 填入你的 LLM API Key
|
||
|
||
# 4. 启动
|
||
python launch.pyw
|
||
```
|
||
|
||
**同样可在 Android 上运行** — 已在 Termux 上测试通过，通过 `python agentmain.py`（CLI 前端）启动：
|
||
|
||
```bash
|
||
# 在 Termux 中
|
||
cd /sdcard/ga
|
||
python agentmain.py
|
||
```
|
||
|
||
启动后告诉 Agent："执行 web setup SOP 解锁浏览器工具"——剩下的它自己搞定。完整引导流程见 [WELCOME_NEW_USER.md](WELCOME_NEW_USER.md)。
|
||
|
||
## 对比
|
||
|
||
| | GenericAgent | OpenClaw | Claude Code |
|
||
|---|---|---|---|
|
||
| 代码量 | ~3,300 行 | ~530,000 行 | 已开源（体量大） |
|
||
| 部署 | `pip install` + API key | 多服务编排 | CLI + 订阅 |
|
||
| 浏览器 | 注入真实浏览器（保留登录态） | 沙箱/无头浏览器 | 通过 MCP 插件 |
|
||
| OS 控制 | 键鼠、视觉、ADB | 多 Agent 委派 | 文件 + 终端 |
|
||
| 自我进化 | 自主生长 SOP 和工具 | 插件生态 | 会话间无状态 |
|
||
| 出厂配置 | 10 个 .py + 5 个 SOP | 数百模块 | 丰富 CLI 工具集 |
|
||
|
||
## 工作原理
|
||
|
||
Agent 拥有 7 个原子工具：`code_run`（执行任意代码）、`file_read/write/patch`（文件操作）、`web_scan`（网页感知）、`web_execute_js`（浏览器控制）、`ask_user`（人机协作）。
|
||
|
||
通过 `code_run`，它可以安装任何包、编写任何脚本、对接任何硬件——相当于在运行时制造新工具。学到的流程保存为 SOP，下次直接调用。
|
||
|
||
核心循环只有 92 行（`agent_loop.py`）：感知 → 思考 → 行动 → 记忆。
|
||
|
||
<details>
|
||
<summary>出厂清单</summary>
|
||
|
||
**核心引擎**：
|
||
- `agent_loop.py` — 感知-思考-行动循环（92 行）
|
||
- `ga.py` — 工具定义与执行
|
||
- `sidercall.py` — LLM 通信（多后端）
|
||
- `agentmain.py` — 会话编排
|
||
|
||
**交互界面**：
|
||
- `stapp.py` — Streamlit Web UI
|
||
- `tgapp.py` — Telegram 机器人
|
||
- `launch.pyw` — 一键启动 + 悬浮窗
|
||
|
||
**基础设施**：
|
||
- `TMWebDriver.py` — 浏览器注入桥接（非 Selenium，通过 Tampermonkey 注入真实浏览器）
|
||
- `simphtml.py` — HTML→文本清洗
|
||
|
||
**5 个核心 SOP**（出厂自带，版本控制）：
|
||
1. `memory_management_sop` — L0 宪法：Agent 如何管理自身记忆
|
||
2. `autonomous_operation_sop` — 自主任务执行
|
||
3. `scheduled_task_sop` — 定时任务
|
||
4. `web_setup_sop` — 浏览器环境引导
|
||
5. `ljqCtrl_sop` — 桌面物理控制（键鼠、DPI 感知）
|
||
|
||
其余一切——Gmail、微信自动化、视觉 API、游戏下载、股票分析——都是 Agent 在使用中自主构建并记忆的。
|
||
|
||
</details>
|
||
|
||
## 许可
|
||
|
||
MIT |