Claude Code, Beyond the Prompt — Hardening an MCP Database Tool (Part 4 Deep Dive)

Short summary

This deep-dive covers production hardening of an MCP database tool for Claude Code, replacing declared rules with enforced ones across seven security layers. Key insights include using structurally read-only database roles (not reversible transaction flags), per-tool scoped identities, purpose-built views instead of base tables, statement timeouts, and row caps. The core principle: a prompt saying 'don't write' is a declaration, while a role without write privileges is enforcement — and the difference matters when models generate unexpected queries.

• •Replace declared rules (prompts, flags) with enforced rules (database privileges, timeouts)
• •Use structurally read-only roles instead of reversible transaction flags for write protection
• •Scope reads per-tool with dedicated roles and purpose-built views to prevent data overexposure

Generated with AI, which can make mistakes.