local-env-mcp
Exposes your local machine's filesystem, git, shell, network, databases, and system to any MCP-compatible LLM client over HTTP.
README
local-env-mcp
A stateless, Dockerized Model Context Protocol (MCP) server built with TypeScript. It exposes your local machine — filesystem, git, shell, network, databases, and system — to any MCP-compatible LLM client over HTTP.
Built natively in TypeScript on the @modelcontextprotocol/sdk utilizing Express, Zod schemas, and a fully decoupled, layered architecture.
Quick Start
# 1. Clone and install
git clone <repo-url>
cd local-mcp
npm install
# 2. Build the TypeScript source
npm run build
# 3. Start the server (runs on port 3000 by default)
npm start
# OR Start with Docker (multi-stage build)
docker compose up -d --build
# 4. Expose via ngrok (in a separate terminal)
ngrok http 3001
# 5. Copy the ngrok URL and add to Claude.ai
# Settings → Integrations → Add MCP Server
# URL: https://your-ngrok-url.ngrok-free.app/mcp
Table of Contents
Architecture
LLM Client (Claude / Cursor / API)
│ HTTPS POST /mcp
▼
Express Routing Layer (src/app.ts)
│
├── GET /health → System Checks
├── POST /mcp → Create Server Session
├── GET /mcp → Resume Session
└── DELETE /mcp → Purge Session
│
▼
Layered Tool Registry (src/tools/registry.ts)
│
┌─────────┬───┴─────┬─────────┬─────────┬─────────┐
▼ ▼ ▼ ▼ ▼ ▼
FS Git Network Shell System Postgres
(9) (11) (8) (9) (8) (3)
Stateless Mode: Each POST request spins up a fresh McpServer instance dynamically managed by a singleton Session Manager.
Project Structure
Migrated from legacy monolithic JS, the project now follows a strictly typed layered framework:
local-mcp/
├── src/
│ ├── server.ts # Listener & bootstrap
│ ├── app.ts # Express routing & middleware wiring
│ ├── config.ts # Zod-validated env loader (.env handled via dotenv)
│ ├── logger.ts # Winston configuration
│ ├── routes/ # Separated Controllers (health.ts, mcp.ts)
│ ├── services/ # Core Logic (processManager, sessionManager)
│ ├── types/ # Shared TS definitions
│ └── tools/ # All 48 MCP Tools separated by domain
│ ├── filesystem/
│ ├── git/
│ ├── network/
│ ├── postgres/ # Dynamic Database capabilities
│ ├── shell/
│ ├── system/
│ └── registry.ts # Injects toolsets into active MCP Sessions
Transport & Session Model
Uses the Streamable HTTP transport.
| Route | Purpose |
|---|---|
POST /mcp |
Initiates tool requests (tools/list, tools/call) — hooks into new or cached session. |
GET /mcp |
Streams Server-Sent Events (SSE) for continuous clients. |
DELETE /mcp |
Gracefully terminate an active server. |
Tool Reference
Currently bundling 48 Native Tools mapped through specific modules.
📁 Filesystem
(All paths resolved relative to FS_ROOT protecting against directory traversal)
read_file,write_file,list_dir,make_dir,delete_path,copy_path,move_path,search_files,file_info
🌿 Git
(Powered by simple-git, targets directories relative to FS_ROOT)
git_status,git_diff,git_log,git_branch,git_add,git_commit,git_push,git_pull,git_clone,git_stash,git_show
🌐 Network
http_request,ping,dns_lookup,port_scan,whois,traceroute,download_file,check_connectivity
💻 Shell
(Managed by the injected ProcessManager service isolating spawn logs in memory)
run_command,spawn_process,list_processes,get_process_logs,kill_process,system_info,get_env,which,ps
🖥️ System
cron_list,disk_usage,open,notify,clipboard_write,clipboard_read,screenshot,list_installed
🐘 Postgres
(Connects dynamically on-the-fly rather than hardcoded global connections)
pg_query: Execute parameterized raw SQL queries directly.pg_list_tables: Enumerate all active tables ignoring system schemas.pg_describe_table: Interrogate type formats, nullability, and default requirements for a target schema table.
Note: The Postgres tools require you to provide a connectionString argument directly through the LLM interface, keeping your server completely credential-less by design.
Environment Variables
Handled seamlessly using dotenv.
| Variable | Required | Default | Description |
|---|---|---|---|
PORT |
No | 3000 |
HTTP port the Express server binds to. |
FS_ROOT |
No | /host-home |
Root directory for File operations. Path constraints enforce this border. |
MCP_AUTH_TOKEN |
Yes (prod) | (none) | Bearer auth token for X-MCP-Token. Omitting heavily warns during dev mode. |
NODE_ENV |
No | development |
Setting to production enforces Token Auth checks. |
Extending with New Tools
Adding tools is significantly easier with the Modular Architecture:
- Create a strictly-typed configuration inside your target domain layer (e.g.
src/tools/custom/myTool.ts):
import { z } from "zod";
import type { ToolDefinition } from "../../types/index.js";
import { ok } from "../../utils/response.js";
export const myCustomTool: ToolDefinition = {
name: "custom_tool",
description: "Does something specific",
schema: {
target: z.string().describe("What to target"),
},
handler: async ({ target }) => {
return ok(`Executed logic on ${target}`);
}
};
- Import and append your tool array to
registerAllToolsinsrc/tools/registry.ts.
Security Model
| Concern | Mitigation |
|---|---|
| Path traversal | safePath() heavily anchors against FS_ROOT bounds. |
| Secret leakage | get_env redacts items masking standard keys (secrets/passwords/apis). |
| Docker Isolation | Runs natively as unescalated mcpuser relying on safe Mount Points. |
| DB Access | Handled dynamically avoiding persistent env leaks in logs. |
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。