supabase-security-mcp

The only Supabase security tool that closes the loop in your AI agent. Audit, preview the fix, and apply it — without leaving Claude / Cursor / Cline.

You: audit my supabase project rkmrsefraqssuyuniyco

Claude: Found 17 critical leaks. Want me to apply all SQL fixes?
        Run preview_fix on each first?

You: preview them all, then apply if safe.

Claude: [previews each]
        All 17 fixes preview cleanly. Applying...
        Done. Re-audited: 0 critical findings remaining.

Other Supabase scanners (SupaExplorer, AuditYourApp, Vibe App Scanner) report. None of them remediate. This one does.

Tools

Install

Claude Desktop / Claude Code

Add to your MCP config (~/.claude.json or claude_desktop_config.json):

{
  "mcpServers": {
    "supabase-security": {
      "command": "npx",
      "args": ["-y", "@perufitlife/supabase-security-mcp@latest"],
      "env": {
        "SUPABASE_ACCESS_TOKEN": "sbp_your_personal_access_token"
      }
    }
  }
}

Get a token at https://supabase.com/dashboard/account/tokens (read+write to your projects).

Cursor

Settings → MCP → Add new MCP server → paste the same JSON object as above.

Cline / Continue / etc.

Anything that supports MCP stdio servers will work — point it at npx -y @perufitlife/supabase-security-mcp@latest with the env var set.

Safety model

• Never auto-applies. Every apply_fix and apply_all_fixes call requires confirm: true.
• Preview before apply. preview_fix runs the SQL inside BEGIN; ... ROLLBACK; so you see if it would error before touching state.
• All-or-nothing bulk apply. apply_all_fixes runs everything in a single transaction. If any statement fails, the entire change rolls back.
• Re-audit after apply. Every apply_fix re-runs the audit and reports whether the finding is actually gone — protects against fix-that-doesn't-fix.
• Read-only by default for audit_project. Token can be a read-only PAT if you only want to scan, never remediate. (For apply_fix, you need write access.)

What it scans

Inherits all checks from supabase-security-skill:

• Tables with RLS disabled and direct anon grants
• SECURITY DEFINER functions executable by anon
• Public storage buckets
• Default privileges still granting CRUD to anon
• Auth signups with autoconfirm enabled
• Defense-in-depth: RLS-locked tables with stale anon grants

Why MCP and not just a CLI

CLIs are great. They're not in your AI agent's context. When you're vibing with Claude in your IDE, asking "is my supabase tight?" should get an actual scan, not a "you should run this command on your laptop."

The flow this enables:

1. "Add a subscriptions table to my schema." → agent does it
2. "Now scan for security issues." → agent calls audit_project
3. Agent notices the new table has no RLS.
4. "Want to fix?" → preview_fix, then apply_fix after you say yes.

That round-trip is the actual product. The audit is just step 2.

Roadmap

• Cron job audit (pg_cron)
• Edge function secrets scan
• Storage object-level RLS scan
• HTML report generation as a tool (returns base64 + filename)
• CORS configuration check

License

MIT.