Short version
MCPAudit Config Review processes pasted MCP client config text for one request. Application code is designed not to store pasted configs, return raw config text, launch configured MCP servers, or contact configured MCP endpoints.
What the scanner receives
When you use the hosted page or ChatGPT app, you provide MCP client configuration text. The scanner parses that text to produce a config-only summary, risk grade, config-health findings, generic server labels in ChatGPT, permission categories, limitations, and next-step guidance.
What the app does not do
- It does not store pasted config text by application design.
- It does not return raw config text to the model or widget.
- It does not launch configured MCP servers.
- It does not contact configured MCP endpoints.
- It does not run connected prompt-injection, SSRF, schema-drift, or SARIF checks inside ChatGPT.
Operational logs
Production logging should stay limited to coarse request status, latency, error class, and non-sensitive aggregate counts. Request bodies, raw configs, env values, endpoint URLs, command argument paths, and host identity should not be logged by application code.
ChatGPT boundary
If you use the ChatGPT app, pasted config text may be part of the ChatGPT conversation and tool-call flow. In Developer Mode, ChatGPT request inspectors can display tool input arguments. Use synthetic configs for review prompts, screenshots, and public examples.
Private configs
For proprietary or secret-bearing configs, prefer the local CLI instead of a hosted or ChatGPT flow:
uvx --from mcp-audits mcp-audit scan
Limits
Config-only review is not a complete MCP security audit and does not certify a server as safe. It distinguishes detected configuration signals from inferred risk. Live tool and resource checks require a connected local scan.