mini_nanobot

Mini NanoBot **Mini NanoBot** is a lightweight **LLM Agent framework** implemented in Python for exploring and practicing the core architectural design of modern AI Agents. The project implements a complete Agent system, including: * Large model invocation * Tool Calling * Skill plugin system * MCP (Model Context Protocol) tool access * Memory * Multi-channel input (CLI / Feishu) The goal of this project is to **implement a clear, extensible Agent architecture with minimal code** to help understand the core working mechanisms of AI Agents. --- # Project Features **Modular Architecture** Decompose the Agent into: * Provider (model provider) * Tool (tool system) * Skill (plugin capability) * MCP (remote tool) * Memory (memory system) * Channel (interaction channel) Each module is designed independently for easy expansion and maintenance. --- **Skill Plugin System** Use ``` skill.yaml + run.py ``` to quickly extend Agent capabilities without modifying the core code. --- **MCP Support** Supports access to remote tool servers via **MCP (Model Context Protocol)** to achieve: * Dynamically load tools * Tool service decoupling * Remote capability extension --- **Multi-Channel Support** Currently supported: * CLI command line chat * Feishu Robot Future scalable: * Web * Slack * Telegram --- # System Architecture The overall system architecture is as follows: ``` 用户 │ ▼ Channel层 (CLI / Feishu Bot) │ ▼ Agent │ ┌──────────┼──────────┐ │ │ │ ▼ ▼ ▼ Provider Tools Memory (LLM) (工具系统) (记忆系统) │ │ ▼ Skill Plugins (YAML + Python) │ ▼ MCP Manager │ ▼ MCP Server（远程工具） ``` The Agent, as the core scheduling module, is responsible for: * Manage conversations * Call the model * Trigger tools * Process tool results * Integrate memory --- # Agent Workflow A complete run of the Agent is as follows: ``` 用户输入 │ ▼ Channel 接收消息 │ ▼ Agent 生成 messages (system prompt + memory + history) │ ▼ 调用 LLM Provider │ ▼ 模型返回响应 │ ├── 普通文本 → 直接返回用户 │ └── Tool Call │ ▼ ToolRegistry 查找工具 │ ▼ 执行 Tool.run() │ ▼ 将 Tool 结果写入 session │ ▼ 再次调用 LLM ``` This loop will continue to execute until: * The model returns the final answer * Or the maximum number of iterations is reached --- # MCP Tool Call Flow MCP (Model Context Protocol) is used to access remote tool servers. Workflow: ``` Agent │ │ 请求工具列表 ▼ MCP Client │ ▼ MCP Server │ │ 返回 tool schemas ▼ Agent ToolRegistry │ │ 注册 MCPTool ▼ LLM 触发工具调用 │ ▼ MCP Client │ ▼ 远程 MCP Server │ ▼ 返回工具执行结果 │ ▼ Agent 继续推理 ``` This design allows: * Tool service decoupling * Multi-Agent shared tools * Remote capability extension --- # Project Directory Structure ``` mini_nanobot/ │ ├── core/ │ ├── agent.py # Agent 核心逻辑 │ ├── response.py # LLM 响应封装 │ ├── prompt_loader.py # Prompt 模板加载 │ └── message.py # Message / Session │ ├── channels/ │ ├── base.py # Channel 接口 │ ├── cli.py # CLI 聊天 │ └── feishu.py # 飞书机器人 │ ├── providers/ │ ├── base.py # Provider 抽象类 │ └── zhipu.py # 智谱AI实现 │ ├── tools/ │ ├── base.py # Tool 抽象类 │ ├── function_tool.py # Function Tool │ ├── registry.py # Tool 注册中心 │ └── builtin/ # 内置工具 │ ├── skills/ │ ├── loader.py # Skill 加载器 │ ├── gold/ # 示例 Skill │ └── weather/ # 天气 Skill │ ├── skill.yaml │ └── run.py │ ├── MCP/ │ ├── client.py # MCP Client │ ├── tool.py # MCP Tool 封装 │ └── manager.py # MCP 管理器 │ ├── memory/ │ ├── base.py # Memory 抽象 │ └── store.py # 简单 Memory Store │ ├── templates/ │ └── mcp_server.py # 示例 MCP Server │ └── main.py # 项目入口 ``` --- # Quick Start ## 1 Install Dependencies ``` pip install -r requirements.txt ``` --- ## 2 Configure API Key For example, using Zhipu: ``` export ZHIPU_API_KEY=your_key ``` --- ## 3 Start CLI ``` python main.py ``` Example: ``` User: 北京天气怎么样？ Agent: 我帮你查询一下... [调用 weather 工具] Agent: 北京今天晴天，气温 26℃。 ``` --- # Extend Capabilities ## Add a New Tool 1. Inherit `Tool` 2. Implement `run()` 3. Register to `ToolRegistry` --- ## Add a New Skill Create a directory: ``` skills/my_skill ``` Add: ``` skill.yaml run.py ``` to load automatically. --- ## Add a New Provider Implement: ``` providers/base.py ``` to support new model services. --- ## Access the MCP Tool Server Add the server address in `MCP Manager` to dynamically load remote tools. --- # Design Goals Mini NanoBot is mainly used to showcase the **core architectural design of AI Agents**: * Module decoupling * Plug-in capabilities * Tool ecosystem * Scalability The code should be kept as **simple, clear, and easy to understand** as possible for easy learning and extension. --- # License MIT License ---