unplug-mcp 0.1.1

MCP server for Unplug — LLM defense layer for Claude Code, Cursor, and more

Unplug MCP

Model Context Protocol server for Unplug — LLM defense layer.

Integrates with Claude Code, Cursor, Windsurf, and any MCP-compatible client.

Installation

pip install unplug-mcp "unplug-ai>=0.3.0"

Optional ML span scanner:

pip install unplug-mcp "unplug-ai[ml]>=0.3.0"

Usage

Local mode (default)

Add to your MCP client configuration:

Cursor — .cursor/mcp.json or Settings → MCP:

{
  "mcpServers": {
    "unplug": {
      "command": "unplug-mcp",
      "args": []
    }
  }
}

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "unplug": {
      "command": "unplug-mcp",
      "args": []
    }
  }
}

Hosted server mode

Point at your Unplug API (same wire format as Guard(mode="server")):

{
  "mcpServers": {
    "unplug": {
      "command": "unplug-mcp",
      "env": {
        "UNPLUG_MODE": "server",
        "UNPLUG_SERVER_URL": "https://api.example.com",
        "UNPLUG_API_KEY": "up_live_xxx"
      }
    }
  }
}

Tools

Tool	Purpose
scan_text	Scan user or retrieved content for injection/leakage
scan_tool_result	Scan tool output before the agent reads it
check_destructive	Gate side-effect tool calls
wrap_untrusted_content	Boundary markers + scan for RAG/web content
session_status	Session taint state for agent hardening

CI

• ci.yml — lint + pytest against PyPI unplug-ai
• pr-scan.yml — regex Guard scan on changed agent/MCP config files (via unplug-scan-pr)

Development

uv sync --extra dev
uv run pytest -q
uv run unplug-mcp

Local SDK path override (monorepo): tool.uv.sources in pyproject.toml.