
一个用Python(FastMCP)编写的沙盒。它在一个工作区文件夹里提供了一个AI代理安全文件、搜索、git-lite和handoff工具,并且通过一个......在会话间作为共享内存。
概览
一个用 Python(FastMCP)编写的沙盒。它在一个工作区文件夹中提供了 AI 代理安全文件、搜索、git-lite 和切换工具,并通过切换协议(PROJECT_STATE、PLAN_LOG、检查点、决策)作为会话间的共享内存。 课程项目:期中考试第一部分 - 自己构建(模型情境协议)。 - 每条路径都与工作区解析; ..\ 遍历、绝对路径和前缀技巧(MCP-activity-evil)被拒绝。- : 文件操作、搜索、git-lite 和 handoff 工具(下表)。- : stdio(Claude 桌面)和可流式 HTTP。- ,由 Windows 服务(NSSM)维持。- 用于 HTTP 端点(MCP_TOKEN)。- (if_version) 确保一个会话无法覆盖另一个会话的更改。- 这样新会话可以仅从文件中恢复工作。 - Windows 10/11 - Python 3.10+,包含mcp包:pip install mcp - Windows版Git(适用于git-lite工具)- Claude Desktop(MCP客户端)- HTTPS部分:cloudflare和NSSM(winget安装Cloudflare)。
概述(AI 翻译)
一个用 Python(FastMCP)编写的沙箱 文件系统 MCP 服务器。它为 AI 代理提供安全文件、搜索、git-lite 和切换工具,集中在一个工作区文件夹内,并通过切换协议(PROJECT_STATE、PLAN_LOG、CHECKPOINTS、DECISIONS)作为会话间共享内存。
课程项目:期中考试第一部分 - 自己构建(模型情境协议)。
特色
- 沙箱工作区根节点。 每条路径都针对工作区解析;[
..\遍历、绝对路径和前缀技巧(MCP-activity-evil)被拒绝。 - 14 个工具:文件操作、搜索、Git Lite 和 handoff 工具(下表)。
- 两种传输方式:stdio(Claude Desktop)和可流式 HTTP。
- 通过Cloudflare隧道实现HTTPS,由Windows服务(NSSM)维持。
- 令牌认证用于HTTP端点(
MCP_TOKEN)。 - 版本锁定(
if_version),使得一个会话无法覆盖另一个会话的更改。 - 切换协议,这样新会话可以仅从文件中恢复工作。
要求
- Windows 10/11
- Python 3.10+,带有
mcp包:pip install mcp - Windows版Git(适用于git-lite工具)
- Claude 桌面(MCP 客户端)
本页由 AI 翻译,技术细节请以原文为准。
README
以下为原文 README 节选,完整内容请查看原文。
StartFilesystemMCP
A sandboxed filesystem MCP server written in Python (FastMCP). It gives an AI agent safe file, search, git-lite and handoff tools inside one workspace folder, and doubles as shared memory between sessions through a Handoff Protocol (PROJECT_STATE, PLAN_LOG, CHECKPOINTS, DECISIONS).
Course project: Midterms Pt 1 - Build Your Own (Model Context Protocol).
Features
- Sandboxed workspace root. Every path is resolved against the workspace;
..\traversal, absolute paths and prefix tricks (MCP-activity-evil) are rejected. - 14 tools: file ops, search, git-lite and handoff tools (table below).
- Two transports: stdio (Claude Desktop) and streamable HTTP.
- HTTPS through a Cloudflare tunnel, kept alive by Windows services (NSSM).
- Token auth for the HTTP endpoint (
MCP_TOKEN). - Version locking (
if_version) so one session cannot overwrite another’s changes. - Handoff Protocol so a new session can resume work from the files alone.
Requirements
- Windows 10/11
- Python 3.10+ with the
mcppackage:pip install mcp - Git for Windows (for the git-lite tools)
- Claude Desktop (MCP client)
- For the HTTPS part:
cloudflaredand NSSM (winget install Cloudflare.cloudflaredandwinget install NSSM.NSSM)
Setup
- Clone the repo into the workspace folder, or copy
mcp_server.pythere. - Set
WORKSPACE_ROOTat the top ofmcp_server.pyto your folder (default:C:\Users\Colee\MCP-activity). - Make sure the four handoff files exist in that folder.
Run over stdio (Claude Desktop)
Add this to claude_desktop_config.json and restart Claude Desktop:
{
"mcpServers": {
"filesystem": {
"command": "C:\\PYTHON314\\PYTHON.EXE",
"args": ["C:\\Users\\Colee\\MCP-activity\\mcp_server.py"]
}
}
}
Run over HTTP
$env:MCP_TOKEN = -join ((48..57)+(97..122) | Get-Random -Count 24 | % {[char]$_})
$env:MCP_TOKEN # copy this value
python mcp_server.py --http # serves http://127.0.0.1:8000/mcp
Without a token the endpoint is open to anyone who has the URL, so always set one.
Expose it over HTTPS (Cloudflare quick tunnel)
cloudflared tunnel --url http://127.0.0.1:8000
The command prints an address like https://.trycloudflare.com. In Claude Desktop add a custom connector named filesystem-http with the URL:
https://.trycloudflare.com/mcp?token=
Quick tunnel addresses change every time the tunnel restarts.
Keep it alive (NSSM, run in an Administrator PowerShell)
nssm install McpServer "C:\PYTHON314\PYTHON.EXE" "C:\Users\Colee\MCP-activity\mcp_server.py --http"
nssm set McpServer AppDirectory "C:\Users\Colee\MCP-activity"
nssm set McpServer AppEnvironmentExtra "MCP_TOKEN=$env:MCP_TOKEN"
nssm set McpServer Start SERVICE_AUTO_START
nssm start McpServer
$cf = (Get-Command cloudflared).Source
安装
This server does not publish a one-line install command.
Open the repository installation guide 新手?先看 mcpskill.org 教程 →配置
{
"mcpServers": {
"filesystem": {
"command": "C:\\PYTHON314\\PYTHON.EXE",
"args": ["C:\\Users\\Colee\\MCP-activity\\mcp_server.py"]
}
}
}