Schema Guard

A Model Context Protocol server that gives an AI agent (Claude Desktop, Claude Code, or any MCP client) tools to audit Postgres / Supabase SQL for security issues — mid-conversation, with structured results. Point it at a schema and it flags missing RLS, permissive policies, broad grants, SECURITY DEFINER search_path gaps, and secrets in client-readable tables.

Built to show MCP / agent tool-use, and tuned to the exact mistakes that no-code exports ship with.

Tools it exposes

Tool	What it does
audit_sql	Full audit of a SQL schema → ranked findings (severity, rule, line, fix) + verdict.
rls_coverage	Quick per-table yes/no of whether RLS is enabled.
list_rules	The checks it runs, with rationale.

Checks

RLS-001 table without RLS · RLS-002 RLS on but no policy · POL-001 USING (true) policy · GRANT-001 broad grant to anon/public · SECDEF-001 SECURITY DEFINER without pinned search_path · COL-001 sensitive column in a client-readable table.

Run the live demo

npm install
npm run demo    # spawns the server, does the MCP handshake, audits a sample schema
npm test        # 10 assertions on the audit engine

The demo is a real MCP client (@modelcontextprotocol/sdk) speaking to the server over stdio — the same handshake Claude performs. It finds 7 issues (2 critical) in fixtures/vulnerable-schema.sql and prints the fix for each.

Wire it into Claude

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "schema-guard": {
      "command": "node",
      "args": ["C:/path/to/35-schema-guard-mcp/src/server.js"]
    }
  }
}

Claude Code — .mcp.json in your project:

{
  "mcpServers": {
    "schema-guard": { "command": "node", "args": ["./src/server.js"] }
  }
}

Then just ask Claude: “audit this schema for RLS gaps” and paste your SQL — it calls audit_sql and reports back.

Stack

Node (ESM) · @modelcontextprotocol/sdk (stdio transport) · zod · zero external services, deterministic output.