ironcurtain

IronCurtain is an early-stage research project aimed at making AI agents safe and genuinely useful. It addresses the challenge of ambient authority, where agents have the same privileges as users, potentially leading to security risks such as data exfiltration or unauthorized actions. By implementing a policy engine that auto-approves actions based on trusted user input, IronCurtain allows agents to perform tasks like managing files and executing commands securely. The framework provides a balance between usability and security, enabling autonomous agents to interact with APIs and manage files while ensuring that their actions are governed by clear policies derived from a constitution-like format. This approach aims to mitigate risks associated with traditional agent frameworks, which often compromise security for functionality.

{
  "mcpServers": {
    "ironcurtain": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-ironcurtain"
      ]
    }
  }
}