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)