Plug AgentGuard into any MCP client — Codex CLI, Claude, or ChatGPT developer mode. Every agent action gets judged before it executes.

{
  "server": {
    "command": "python",
    "args": ["-m", "app.mcp_server"],
    "cwd": "backend"
  },
  "client": {
    "transport": "streamable-http",
    "url": "http://127.0.0.1:8001/mcp"
  }
}

• declare_missionDeclare or replace an agent mission and include it in future audit snapshots.
• request_actionRequest authorization before any high-risk action; proceed only when status is allowed.
• check_action_statusPoll a pending action after the human approval decision.
• get_policiesRead the active caps, thresholds, and allowlists for an agent.