Tool & Permission Policies
Tool Policies represent the critical security boundary between AI intention and real-world execution. MUJPI enforces strict, declarative policies that determine exactly which tools an agent can use, under what conditions, and when explicit human approval is required.
Architecture & Data Flow
Every tool invocation attempt is intercepted by the Policy Engine, which evaluates it against defined rules before allowing execution or suspending the task for approval.
Data Flow
Tool payloads are held in memory during suspension. If approved, the payload is forwarded. If rejected, the agent receives a policy violation error and the event is audited.
Typical Use Cases
- ✦Requiring a manager's approval before an agent executes a server reboot
- ✦Limiting database read access to specific schemas based on the user's role
- ✦Blocking an agent from sending external emails outside business hours
- ✦Enforcing expenditure limits on automated purchasing tools
Capabilities
- ✦Role-based access control (RBAC) for agents
- ✦Human-in-the-loop (HITL) interception
- ✦Time-based and contextual execution restrictions
- ✦Automated policy violation alerting
Security Considerations
- Fail-secure architecture: unmapped or unrecognized tools are blocked by default.
- Human approvals are cryptographically signed to prevent spoofing.
- Policy definitions are version-controlled and immutable during execution.
Setup & Deployment
Deployment Modes
Setup Overview
Write policy definitions using the visual rule builder or YAML. Define approval routing (e.g., notify Slack channel X for approval) for sensitive actions.