If you're building agents that call external APIs or trigger risky operations, this server sits between the LLM and execution to create an approval layer. It exposes five read-only tools like validate_action_card and resolve_action_card that return Decision Receipts and trace references instead of running the action directly. Your code keeps control of actual execution while the agent gets cryptographic proof that a human or policy engine approved the step. Requires a Neura Registry Agent Passport for production use. Auth is bearer token over streamable HTTP, and all errors return safe references you can share in support threads without leaking secrets.
claude mcp add --transport http com.neurarelay-relay-mcp https://www.neurarelay.com/mcp