inspector

Least-privilege permissions

• Restrict user access: Only grant "inspector" access to team members who actively debug MCP servers; remove access when roles change.
• Separate credentials: Never use production API keys or passwords in inspector. Create test-only credentials with minimal permissions for evaluation purposes.
• Audit log access: Limit who can view or export inspection logs, since they may contain sensitive data or authentication tokens.
• Environment segregation: Run inspector in a separate, isolated environment from production systems where real customer data flows.

Network isolation

• Air-gap testing: When possible, run inspector on machines disconnected from production networks to prevent accidental connections to live systems.
• Firewall rules: Block inspector instances from accessing production databases, APIs, or internal services; only allow connections to designated test servers.
• VPN containment: If inspector must connect to remote servers, use a separate VPN tunnel dedicated to testing—never share production VPN credentials.
• Egress filtering: Monitor and restrict outbound connections from inspector to prevent data leakage to unauthorized external services.

Monitoring

• Log retention limits: Set automatic deletion of inspector logs after 7–30 days to prevent long-term accumulation of sensitive debugging data.
• Access logging: Track who accessed inspector, when, and what they inspected; flag unusual access patterns or after-hours activity.
• Secret scanning: Regularly scan inspector logs and saved test files for accidentally committed API keys, tokens, or passwords using automated tools.
• Update checks: Subscribe to security alerts for the @mcpjam/sdk and CLI packages; apply patches within 72 hours of release.

Red flags to watch for

• Credentials in chat: If passwords, API keys, or tokens appear in inspection chat or outputs, immediately rotate those credentials and review what was logged.
• Production connections: Alert if inspector connects to real production systems or databases instead of test environments—this indicates misconfiguration.
• Unexpected data exports: Bulk downloads of inspection results or conversations without documented business reason may signal unauthorized data gathering.
• Bypassed safety checks: If you observe inspector revealing or circumventing safety guardrails designed to protect connected AI agents, stop use and escalate to security team.