Agent Cage

What it does

Agent Cage provides a secure, sandboxed Ubuntu environment for AI agents to operate within. It separates the agent's capabilities from its network access, allowing operators to enforce strict security policies using professional IT tools like proxy policies and TLS inspection.

Tools

• agentcage up: Starts the sandbox environment with optional MCP support.
• agentcage shell: Provides a terminal interface to the sandboxed agent.
• agentcage dashboard: Launches a web-based control panel for monitoring and policy management.
• agentcage audit: Filters and views blocked network requests for security analysis.
• agentcage policy: Manages and reloads network access policies.

Installation

To run Agent Cage with MCP support, use the following configuration in your claude_desktop_config.json:

{
  "mcpServers": {
    "agent-cage": {
      "command": "agentcage",
      "args": ["up", "--mcp"]
    }
  }
}

Supported hosts

Confirmed support for Claude Code, Cline, LangGraph, and SSH-based integrations.