Micro-Agent
mcp
Warn
Health Warn
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 7 GitHub stars
Code Pass
- Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
- Permissions — No dangerous permissions requested
Purpose
This tool provides a lightweight, Python-based AI agent framework tailored for vertical domain applications. It uses a ReAct execution loop along with Skills, RAG, and Knowledge Graph components to build specialized, domain-specific agents.
Security Assessment
The overall risk is rated as Medium. The automated code scan found no dangerous patterns, hardcoded secrets, or requests for overly broad system permissions. However, the application is designed to execute tasks dynamically, and the README specifically mentions the inclusion of a "REPL sandbox execution" environment. Any tool capable of executing AI-generated code locally requires strict isolation and oversight. Additionally, the setup process involves handling sensitive LLM API keys, which must be stored securely in the local environment.
Quality Assessment
The project benefits from a clear MIT license and recent development activity, with the last code push occurring today. Its technical architecture and documentation are remarkably comprehensive and well-structured. However, it suffers from very low community visibility, currently sitting at only 7 GitHub stars. This means it has not undergone extensive peer review, and long-term maintenance relies almost entirely on the original creator.
Verdict
Use with caution: the code is currently safe and well-documented, but low community adoption and built-in code execution capabilities mean you should employ strict sandboxing when running it.
This tool provides a lightweight, Python-based AI agent framework tailored for vertical domain applications. It uses a ReAct execution loop along with Skills, RAG, and Knowledge Graph components to build specialized, domain-specific agents.
Security Assessment
The overall risk is rated as Medium. The automated code scan found no dangerous patterns, hardcoded secrets, or requests for overly broad system permissions. However, the application is designed to execute tasks dynamically, and the README specifically mentions the inclusion of a "REPL sandbox execution" environment. Any tool capable of executing AI-generated code locally requires strict isolation and oversight. Additionally, the setup process involves handling sensitive LLM API keys, which must be stored securely in the local environment.
Quality Assessment
The project benefits from a clear MIT license and recent development activity, with the last code push occurring today. Its technical architecture and documentation are remarkably comprehensive and well-structured. However, it suffers from very low community visibility, currently sitting at only 7 GitHub stars. This means it has not undergone extensive peer review, and long-term maintenance relies almost entirely on the original creator.
Verdict
Use with caution: the code is currently safe and well-documented, but low community adoption and built-in code execution capabilities mean you should employ strict sandboxing when running it.
A lightweight AI agent framework for vertical domain applications | 面向垂域应用的轻量级 AI Agent 框架
README.md
为什么选择 Micro-Agent
| Micro-Agent | LangGraph | AutoGen | OpenClaw | |
|---|---|---|---|---|
| 定位 | 垂域 Agent 服务 | 通用工作流编排 | 多智能体协作 | 个人自治助手 |
| 架构 | ReAct Loop + 可插拔组件 | 有向图状态机 | 会话驱动消息传递 | 3 层技能 + 持久自治 |
| 部署 | Docker 一键启动,API 即用 | 需自行搭建服务层 | 需自行搭建服务层 | 独立进程,24/7 常驻 |
| 体量 | 核心 <3000 行,依赖精简 | 依赖 LangChain 全家桶 | 中等,微软维护 | 插件生态庞大但安全隐患多 |
| 模型 | litellm 统一接口,一行切换 | 依赖 LangChain 适配层 | 原生多模型支持 | 多模型支持 |
| 垂域能力 | Skills + RAG + KG 原生支持 | 需自行组装 | 需自行组装 | 通用自动化,非垂域 |
| 适合场景 | 行业知识驱动的专业 Agent | 复杂多步工作流 | 多角色协作 | 个人效率自动化 |
架构
核心组件:
- LLM Layer — 通过 litellm 统一接口,一套代码切换 OpenAI / DeepSeek / Claude / Ollama 等任意模型
- Agent Core — ReAct 执行引擎(Think → Act → Observe 循环),支持 SubAgent 子任务分发与 REPL 沙箱执行
- Memory — 会话记忆系统,支持短期记忆、文件持久化、跨会话恢复
- Skills — 将领域规范、编码标准等知识注入 Agent 的 system prompt,使其具备专业能力
- RAG — 从领域知识库中检索相关文档,为推理提供上下文
- Knowledge Graph — 基于图结构的领域知识表示与关联推理
- MCP / Tools — 通过 Model Context Protocol 连接外部工具和数据源
快速开始
环境要求
- Python ≥ 3.11
- 任意 LLM API Key(OpenAI / DeepSeek / Claude / Ollama / OpenRouter 等)
安装
git clone https://github.com/fdueblab/Micro-Agent.git
cd Micro-Agent
pip install -e ".[dev]"
配置
cp .env.example .env
编辑 .env,填入 API Key:
LLM_MODEL=deepseek/deepseek-chat
LLM_API_KEY=sk-xxx
支持任何 litellm 兼容的模型格式,如
openai/gpt-4o、ollama/qwen2.5、openrouter/qwen/qwen3-coder-flash等。
启动
uvicorn api.app:app --host 0.0.0.0 --port 8010 --reload
访问 http://localhost:8010/docs 查看 API 文档。
Docker 部署
docker-compose up -d
定义垂域任务
只需三步,即可将通用 Agent 转化为面向特定领域的专业智能体:
1. 编写 Prompt 模板
{# task/templates/code_review.md.j2 #}
请对以下代码进行审查,重点关注安全性和性能:
代码路径: {{ code_path }}
审查标准: {{ standards }}
2. 注册任务
# task/builtin.py
register_task(TaskConfig(
name="code_review",
prompt_template="code_review.md.j2",
system_prompt="你是一名资深代码审查工程师。",
llm_profile="reasoning",
max_steps=20,
))
3. 调用
curl -X POST http://localhost:8010/api/tasks \
-H "Content-Type: application/json" \
-d '{"prompt": "审查 src/main.py", "agent_name": "code_review"}'
多 LLM Profile
为不同场景配置不同的模型策略:
# config/config.toml
[llm.default]
model = "deepseek/deepseek-chat"
temperature = 0.0
max_tokens = 8192
[llm.fast]
model = "deepseek/deepseek-chat"
max_tokens = 4096
timeout = 30
[llm.reasoning]
model = "openai/o1-mini"
max_tokens = 16384
timeout = 120
任务中通过 llm_profile 指定:
register_task(TaskConfig(
name="my_task",
llm_profile="reasoning", # 使用推理模型
...
))
内置示例任务
项目内置了多个真实场景的 Agent 任务作为参考实现:
| 任务 | 说明 | 垂域组件 |
|---|---|---|
| 代码分析 | 上传代码 → 自动分析函数结构 | Tools |
| 服务封装 | 上传代码 → 自动生成 Docker + MCP 服务 | Skills + RAG + Memory |
| 算法模型生成 | 描述需求 → 生成算法模型代码 | Skills + RAG + Memory |
| MCP 服务测试 | 连接 MCP 服务器 → 自动发现并测试工具 | MCP |
| 服务评测 | 上传数据 → 自动执行评测并输出报告 | Tools |
| AML 模型评测 | 上传数据 → 多指标安全评测(支持数据适配) | MCP + Tools |
这些任务展示了如何通过组合 Skills、RAG、MCP 等组件,将通用 Agent 打造为垂域专业智能体。你可以参考它们的实现来构建自己的任务。
扩展点
| 组件 | 接口 | 内置实现 | 可扩展方向 |
|---|---|---|---|
| 模型 | litellm | OpenAI, DeepSeek, Claude | Ollama, vLLM, 任意 OpenAI 兼容 API |
| 工具 | Tool ABC |
Bash, MCP, Terminate | 任意自定义工具 |
| 记忆 | MemoryProvider |
ShortTermMemory, FileMemory | Redis, 向量数据库 |
| 检索 | Retriever |
EmbeddingRetriever | FAISS, ChromaDB, Milvus |
| 技能 | Skill + SkillRegistry |
SKILL.md 目录发现 | 远程技能市场 |
项目结构
Micro-Agent/
├── core/ # Agent 核心引擎
│ ├── agent.py # ReAct 循环执行引擎
│ ├── llm.py # LLM 统一调用层 (litellm)
│ ├── config.py # 配置管理 (TOML + 环境变量)
│ ├── memory/ # 记忆系统 (短期 / 持久化)
│ ├── rag/ # 检索增强 (Embedding)
│ ├── skill/ # 技能系统 (注册 / 发现 / 注入)
│ └── schema.py # 数据模型 (Event / Message / ToolCall)
├── tool/ # 工具层
│ ├── base.py # Tool 抽象接口
│ ├── bash.py # Bash 命令执行
│ ├── mcp/ # MCP 工具 (stdio / SSE)
│ └── registry.py # 工具注册表
├── task/ # 任务定义
│ ├── base.py # TaskConfig + 模板渲染
│ ├── builtin.py # 内置任务注册
│ └── templates/ # Jinja2 Prompt 模板
├── api/ # API 服务层
│ ├── app.py # FastAPI 入口
│ ├── routes/ # 路由 (任务管理 / Agent 端点)
│ └── services/ # SSE 流 / 文件处理
├── workspace/ # 工作区
│ ├── knowledge/ # RAG 知识库文档
│ └── skills/ # Skill 定义 (SKILL.md)
├── config/ # 配置文件
├── tests/ # 测试
└── deploy/ # Docker 部署
许可
Reviews (0)
Sign in to leave a review.
Leave a reviewNo results found