基础设施管理MCP服务器集合

MCP-Servers

A collection of Model Context Protocol (MCP) server implementations (some authored in-house, some vendored third-party frameworks/packages), used for AI-assistant integration with home lab and Steamroller Studios infrastructure.

Overview

Security posture

Several of the in-house servers (proxmox-mcp, unifi-mcp, zabbix-mcp, bconsole-mcp) went through a security-hardening pass driven by a CISSP-agent review. The consistent principle applied: make mutating/destructive tools safe by default in code, not by disabling them. Rather than shipping a global read-only toggle as the only safeguard, every destructive or high-impact tool call now requires the caller to pass an explicit confirm=true (or, for the highest-risk actions, two independent flags; see each server's README for its exact gate names). A tool call without confirmation raises an error before any request reaches the underlying API/SSH/subprocess layer. Nothing is silently blocked, but nothing destructive runs by accident either.

Where relevant, additional hardening was layered on top of confirmation gates:

• Input validation: RFC1918/port-range checks on unifi-mcp port forwards; strict ObjectID pattern checks on all UniFi _id params; VMID/identifier sanitization in proxmox-mcp
• Fail-closed TLS: proxmox-mcp requires a second explicit opt-in (PROXMOX_ALLOW_INSECURE_TLS) to actually disable certificate verification, rather than silently downgrading on PROXMOX_VERIFY_SSL=false
• Command-injection protection: bconsole-mcp rejects any caller-controlled value containing a newline before it can reach bconsole's SSH-piped stdin REPL
• Escape-hatch allowlisting: bconsole-mcp's raw console() passthrough is restricted to read-only verbs, closing the loophole where it could otherwise bypass every other tool's confirm gate
• Privilege-escalation dual-gating: zabbix-mcp's user_update and proxmox-mcp's delete_instance require a second, distinct flag beyond confirm=true for their highest-risk case (granting Zabbix super-admin; deleting a VM/container)

See each server's README for its full tool list and exact gate names/parameters.

What is MCP?

The Model Context Protocol (MCP) is an open standard created by Anthropic that provides a universal way to connect AI assistants to external data sources and tools. It uses JSON-RPC 2.0 for communication and supports multiple transport protocols (stdio, HTTP, SSE).

For detailed technical background, see MCP-web-research.md.

Repository

git clone git@github.com:dwpdkp/mcp-servers.git
cd mcp-servers

Each in-house server directory has its own README with setup, environment variables, .mcp.json client config, and a full tool reference. Start there for any given server. fastmcp/ and MCP-Builder-Stuff/ are vendored third-party dependencies, not standalone servers to run directly.

Development

Most servers use uv as the package manager.

cd <server-dir>
uv sync
uv run pytest    # where a test suite exists (proxmox-mcp, unifi-mcp, zabbix-mcp)

See CLAUDE.md for repo-wide development guidelines.

Documentation

• CLAUDE.md: development guidelines for Claude Code working in this repo
• MCP-web-research.md: comprehensive MCP technical reference
• Per-server README.md: setup, tools, and security model for each server

License

• proxmox-mcp, unifi-mcp, bconsole-mcp, bacularis-mcp, weather: MIT
• zabbix-mcp: MIT (fork of mhajder/zabbix-mcp, also MIT)
• unraid-mcp: see upstream project
• mcp_youtube_dlp: GPL-2.0
• fastmcp, MCP-Builder-Stuff: see vendored project license

Resources

• MCP Official Documentation
• MCP Python SDK
• FastMCP Documentation

Document History