claude-agent-orchestrator

The system utilizes Auth0 M2M for authentication with the Coordinator, requiring sensitive credentials (client_secret) to be securely managed via environment variables. The embedded MCP server and Runner Gateway are bound to localhost by default, which is good for restricting network exposure. However, the 'claude-code' executor profile often uses 'bypassPermissions' mode, allowing AI agents extensive system access within their `project_dir` without user prompts. This is a significant privilege and a potential risk if agents or prompts are untrusted, making security highly dependent on the integrity of the agents themselves. Placeholder resolution and environment variable injection into MCP configurations (e.g., in `claude_client.py`) could also be vectors if the runner's environment is compromised.