mcp-server

The server can access confidential data. It directly executes LLM-generated SQL queries via `assetdb_sql_query`, which poses a high risk of SQL injection, especially with 'text' parameters that use template substitution rather than prepared statements. By default, write operations (saving queries, archiving queries, executing GraphQL mutations) are disabled, significantly reducing risk. However, enabling `_ALLOW_` environment variables grants the LLM the same powers as the authenticated user, allowing potentially destructive or data-modifying actions. Full query results are saved locally, potentially exposing sensitive data. There are no hardcoded secrets or 'eval' usage observed. Proper authentication via `stacklet-admin login` is required, and permissions are inherited from the authenticated user. The `json_guard` decorator is used to prevent accidental JSON encoding issues from LLMs, but does not prevent malicious input.