Production-grade Model Context Protocol (MCP) server toolkit for building AI Agents with persistent memory, task scheduling, sandboxed workspace, and web browsing — all in one.
mcp-server-agentic-toolkit
给你的 LLM 一个大脑、双手和记忆。 Give your LLM a Brain, Hands, and Memory.
一个 MCP Server,能把任何兼容 MCP 的 LLM 客户端(Claude Desktop、Cursor、Windsurf 等)变成一个功能完整的 AI Agent —— 支持持久化记忆、代码执行、网页浏览和定时任务调度。
An MCP Server that turns any compatible LLM client (Claude Desktop, Cursor, Windsurf, etc.) into a full-featured AI agent — with persistent memory, code execution, web browsing, and task scheduling.
Architecture / 项目架构
本项目采用三层抽象架构(Three-layer abstraction architecture):
- MCP 接口层 / MCP Layer — 将核心能力封装为 MCP 的 Tools、Resources 和 Prompts
- 核心逻辑层 / Core Layer — 记忆管理、工作区沙箱、浏览器自动化、任务调度
- 存储层 / Storage — 向量数据库、元数据、文件系统
┌──────────────────────────────────────────────────┐
│ LLM Client (Claude / Cursor) │
│ MCP Protocol │
├──────────────────────────────────────────────────┤
│ MCP Layer │
│ ├── Tools: memorize, recall, browse, exec, ... │
│ ├── Resources: memory://short-term, memory://stats│
│ └── Prompts: researcher, coder, assistant │
├──────────────────────────────────────────────────┤
│ Core Layer │
│ ├── MemoryManager (ChromaDB + SQLite) │
│ ├── WorkspaceManager (Docker sandbox + Git) │
│ ├── BrowserManager (Playwright) │
│ └── TaskScheduler (APScheduler) │
├──────────────────────────────────────────────────┤
│ Storage │
│ ├── ChromaDB (vector embeddings) │
│ ├── SQLite (metadata + tasks) │
│ └── Filesystem (workspace sandbox) │
└──────────────────────────────────────────────────┘
Quick Start / 快速开始
Install / 安装
pip install mcp-server-agentic-toolkit
Configure / 配置 MCP 客户端
在客户端的 MCP 配置文件中添加(Add to your client's MCP config, e.g. ~/.claude/claude_desktop_config.json):
{
"mcpServers": {
"agentic-toolkit": {
"command": "python",
"args": ["-m", "mcp_agentic_toolkit"]
}
}
}
首次运行会自动在 ~/.agentic-toolkit/ 创建默认配置。
On first run, the server auto-creates ~/.agentic-toolkit/ with default configuration.
Verify / 验证安装
npx @modelcontextprotocol/inspector python -m mcp_agentic_toolkit
Tools / 工具 (17)
Memory / 记忆工具
| Tool / 工具 | Description / 说明 |
|------|------|
| memorize(content, importance, tags) | Save information to long-term memory / 将信息存入长期记忆 |
| recall(query, top_k) | Semantic search across stored memories / 在已存储的记忆中进行语义搜索 |
| forget(memory_id) | Delete a specific memory / 删除指定记忆 |
| memory_stats() | Get memory store statistics / 获取记忆存储统计信息 |
Browser / 浏览器工具
| Tool / 工具 | Description / 说明 |
|------|------|
| browse(url) | Fetch and extract readable content from a URL / 抓取并提取网页正文内容 |
| screenshot(url) | Take a screenshot (returns base64 PNG) / 截取网页截图 |
| search(query, engine) | Search the web (DuckDuckGo) / 搜索网页 |
Execution / 代码执行工具
| Tool / 工具 | Description / 说明 |
|------|------|
| run_code(code, language) | Execute code in an isolated Docker sandbox / 在 Docker 沙箱中执行代码 |
| run_command(command) | Run a shell command in the workspace / 在工作区执行 Shell 命令 |
| read_file(path) | Read a file from the workspace / 读取工作区文件 |
| write_file(path, content) | Write content to a workspace file / 写入内容到工作区文件 |
| git_status() | Check workspace Git status / 查看工作区 Git 状态 |
| list_files(path) | List files in a workspace directory / 列出工作区目录下的文件 |
Scheduler / 调度工具
| Tool / 工具 | Description / 说明 |
|------|------|
| schedule_task(name, cron, tool_name, arguments) | Schedule a recurring task / 设置定时任务 |
| list_tasks() | List all scheduled tasks / 列出所有已调度的任务 |
| cancel_task(task_id) | Cancel a scheduled task / 取消定时任务 |
| run_task_now(task_id) | Trigger a task immediately / 立即触发指定任务 |
Resources / 资源
| Resource / 资源 | Description / 说明 |
|----------|------|
| memory://short-term | Recent 10 memories as JSON / 最近 10 条记忆 |
| memory://stats | Memory store statistics / 记忆存储统计 |
Preset Workflows / 预设工作流
examples/ 目录提供开箱即用的 YAML 工作流模板(The examples/ directory contains YAML workflow templates):
- Research Assistant / 研究助手 — 自动搜索、阅读、记忆并生成报告 (auto search, read, memorize, and compile reports)
- Code Reviewer / 代码审查 — 编写代码、运行测试、Git 提交 (write code, run tests, commit with Git)
- News Analyzer / 新闻分析 — 浏览新闻、生成摘要、定时推送 (browse news, summarize, schedule daily digests)
Configuration / 配置文件
默认路径 / Default config: ~/.agentic-toolkit/config.yaml
memory:
db_path: "~/.agentic-toolkit/chromadb"
embedding_model: "all-MiniLM-L6-v2"
max_memories: 10000
workspace:
base_dir: "~/.agentic-toolkit/workspace"
docker_image: "python:3.12-slim"
timeout_seconds: 30
max_output_bytes: 10240
browser:
headless: true
timeout_ms: 30000
search_engine: "duckduckgo"
scheduler:
db_path: "~/.agentic-toolkit/tasks.db"
Development / 本地开发
git clone https://github.com/EDY2026/mcp-server-agentic-toolkit.git
cd mcp-server-agentic-toolkit
pip install -e .
pytest tests/ -v
python -m mcp_agentic_toolkit