MCP~85 · IA en attente

I build a Supabase security skill for Claude, and now I built its read-only MCP server (open source)

r/ClaudeCodeu/Farenhytee30 septembre 2026

Capture du projet

Analyse IA en cours de préparation : les informations ci-dessous proviennent de la détection automatique.

Résumé

A few months ago I built Database Sentinel as a Claude Skill. It was a set of instructions, audit queries and fix templates that Claude would follow to check a Supabase (or MongoDB) project for the usual vibe-coding mistakes: RLS off, USING (true) policies, SECURITY DEFINER functions anyone can call through /rpc, serv…

Afficher le post original
A few months ago I built Database Sentinel as a Claude Skill. It was a set of instructions, audit queries and fix templates that Claude would follow to check a Supabase (or MongoDB) project for the usual vibe-coding mistakes: RLS off, `USING (true)` policies, SECURITY DEFINER functions anyone can call through `/rpc`, service-role keys sitting in frontend code. It worked, but it bugged me. A skill is just instructions. It runs with whatever database access Claude already has, and it only works in Claude. So I turned it into an MCP server. Same audit, but now the tools themselves are locked down, and it works in Claude Code, Cursor, Claude Desktop, or any MCP client. What it gives your model: - 20 fixed audit queries (RLS, policies, grants, definer functions, buckets). Nothing else runs. - Schema with anon/authenticated grants and function bodies. No rows. - An anon-key probe: "can a stranger read this table?" It returns a status code and row count, never data. - A repo scan for leaked service-role keys and JWT secrets. It reports file and line, never the value. It connects as its own `sentinel_auditor` role: read-only, 5s timeout, no SELECT on your tables. There are no write or drop tools. Fixes come back as SQL for you to review. Does it actually work? I wrote 25 labeled Supabase projects, and a separate agent wrote 10 of them blind; those are hash-locked as a test set. The audit rules scored F1 0.85 there and caught every critical issue. That's with a single prompt and a small open model. Then I ran it on two of my own hosted projects. It found real problems. It also found a bug in itself: it called every RLS-off table "exposed" even when anon had no grants on it. That's fixed in 1.0. Try it - Claude Code: `claude plugin marketplace add Farenhytee/database-sentinel`, then `claude plugin install database-sentinel@database-sentinel` - Cursor: one-click button in the README - Setup guide: https://github.com/Farenhytee/database-sentinel/blob/main/docs/mcp.md The skill is still there if you prefer it, and it still covers MongoDB. The MCP server is Supabase-only for now. PRs are always welcome. Happy coding.
I build a Supabase security skill for Claude, and now I built its read-only MCP server (open source) · Claude Project Scanner