Frequently Asked Questions (FAQs)

Frequently Asked Questions (FAQs)

📌 TL;DR: Quick answers to common operational and security questions about MCP deployment, rollback, isolation, and logging.

1. How do I restart a stuck MCP server?

If running via Docker, use: docker restart mcp_server If deployed via a systemd service: systemctl restart mcp.service

2. Can I audit historical prompt results?

Yes, if logging.enabled=true in the config. All prompt-response pairs are stored in append-only JSONL or sent to observability platforms (e.g., Loki, Vector.dev).

3. Can I anonymize user inputs automatically?

Yes. MCP supports middleware hooks (pre-processing) that allow regex-based or AI-based redaction of names, IPs, domains, and emails.

4. How do I roll back to a previous model or tool config?

If using Git-based deployment (recommended), use: git checkout For container-based rollbacks, tag images with semantic versioning (e.g., v1.2.3) and use: docker run socradar/mcp:v1.2.2 Finally, to roll back: POST /api/models/rollback { “model_id”: “llm-claude3-secure”, “version”: “v1.6.2” }

5. What if a tool crashes or times out?

Each tool execution is wrapped with timeout and error capture logic. Use timeout_seconds in config and enable fallback_tool for graceful degradation.

6. Can I run multiple MCP flows in parallel?

Yes. MCP is async-first and supports concurrent executions per user/API key, configurable via MAX_CONCURRENT_TASKS.

7. How can I test a new prompt flow without affecting production?

Set the environment flag MCP_ENV=staging and use a separate model registry or API key. Use dry_run=true in execution to simulate.

8. How do I secure prompt inputs and prevent injection?

Use:

• Prompt sanitization middleware
• Schema validation (e.g., JSONSchema)
• Strict input types (no raw shell commands passed)

9. Can I deploy MCP in an air-gapped environment?

Yes, if using self-hosted LLMs and internal tool registry. Disable all outbound connections in config and replace remote APIs with mocks/stubs.

10. What are the recommended system requirements?

• 8+ cores
• 32 GB RAM
• SSD-backed storage
• Optional GPU if using on-prem LLMs (NVIDIA A10 or higher)