MCP connection review — AI Anonymous by JABSystems
Article: https://jabsystems.io/ai-anonymous/mcp-read-the-tool-list

Record server origin/version, review date, host, runtime account and credential scope.
For each tool record: name, required inputs, declared hints, actual implementation, data destination, side effects, approval needs.
READ: Which records or files can it see? Reading can still expose sensitive data.
WRITE: Which records or files can it change? Can changes be reviewed/reversed?
SEND: Can it message people, publish content, or send data to another service?
Check code and runtime permissions; annotations and prompts are not enforcement.
Limit exposed tools and credentials to the task; use host approval controls where supported.
Review all catalog pages and repeat after server or permission changes.
Only exercise known tools with synthetic data in a suitable test environment.
Inspector connects to a server; stdio inspection launches its code. A venv isolates packages, not operating-system permissions.
This checklist is a proposed review workflow, not a security assessment.