The Supabase deploy-gate that runs where you vibe-code — and blocks only on the holes that are truly dangerous.

In 2026, a wave of AI-built ("vibe-coded") apps shipped to production with the same handful of security holes. CVE-2025-48757 and the Lovable disclosure cataloged 170+ live apps leaking user PII — most through one anti-pattern: a Row Level Security policy that lets any logged-in user read everyone's rows. Studies of vibe-coded apps found ~80% repeat the same RLS mistake and ~72% ship a hardcoded secret somewhere.

rlsgate is a local-first, static scanner that reads your Supabase migrations, RLS policies, env files, and frontend bundle and blocks the deploy when it finds the finite set of high-signal holes that actually leak. No database connection, no telemetry, no account. It runs in seconds, as a CLI, a Claude Code skill, or a GitHub Action.

$ rlsgate scan
rlsgate  ·  1 migrations, 1 env, 2 source files scanned

✖ CRITICAL  Table public.profiles holds personal data but Row Level Security is never enabled
    where: supabase/migrations/0001_init.sql:2
    why:   With RLS off in a client-reachable schema, anyone holding the public anon key can read every row.
    fix:   ALTER TABLE public.profiles ENABLE ROW LEVEL SECURITY; then add an ownership policy.

✖ CRITICAL  RLS policy 'read all posts' grants every authenticated user access to public.posts
    where: supabase/migrations/0001_init.sql:15
    fix:   Bind the policy to the row owner, e.g. USING (auth.uid() = user_id).

▲ HIGH      Webhook handler does not verify the request signature (app/api/stripe/webhook/route.ts)
    ...
  3 critical  2 high
$ echo $?
1

Why it's different

Plenty of tools scan vibe-coded apps. rlsgate is the one that is local-first, runs inside your build loop, and blocks at deploy — not a dashboard you visit, not a sidecar you run in production.

• vs hosted SaaS scanners — they guess at your security from outside your live URL and want a token to your repo. rlsgate reads your actual migrations and RLS policies locally. No DB connection, no account, no token handoff, no telemetry.
• vs runtime sidecars / firewalls — they filter requests after deploy, in front of your app. rlsgate catches the hole before deploy, in your repo, where it's a one-line fix instead of an incident.
• vs broad scanners — they compete on rule count (hundreds to thousands), which means noise, which means you turn them off. rlsgate ships a finite, high-precision set of checks for the holes that actually leak — including the CVE-2025-48757 "any authenticated user can read any row" class that breadth scanners miss — so it's a gate you'll actually leave on.

It's the same checks across three surfaces with full parity: a CLI, a Claude Code skill that fires mid-build (inside the AI tool generating your app), and a GitHub Action that blocks the merge. Works across any builder — Lovable, Bolt, Cursor, v0, or hand-rolled.

Install

pip install rlsgate         # or: uv tool install rlsgate
rlsgate scan                # scan the current directory

No Python project? A zero-install run:

uvx rlsgate scan ./my-app

Usage

rlsgate scan [PATH]              # scan a project (default: .)
rlsgate scan --json             # machine-readable output
rlsgate scan --fail-on critical # only CRITICAL fails the gate (default: high)
rlsgate scan --ai               # add AI-drafted fix snippets (needs ANTHROPIC_API_KEY)
rlsgate rules                   # list the checks

Exit codes: 0 clean (or only findings below --fail-on) · 1 blocking findings · 2 usage error. The non-zero exit is what makes it a gate.

What it checks (v1)

The Supabase anon key (a public JWT with role: anon) is intentionally public and is never flagged — only the dangerous service_role key is.

GitHub Action

# .github/workflows/rlsgate.yml
name: rlsgate
on: [pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: GerardoRdz96/rlsgate@v0.1.0
        with:
          path: .
          fail-on: high

Claude Code skill

Copy skill/ into your project's .claude/skills/rlsgate/ (or install the repo as a plugin). Then, in the agent loop before deploy:

The skill runs the gate and brings the findings into the conversation so the agent fixes them before shipping.

What rlsgate is NOT

• Not a full security audit. It checks a finite, high-signal ruleset, statically. It will not catch logic bugs, business-rule auth flaws, or holes that only appear at runtime.
• Not a database scanner. v1 reads files only — it never connects to your Supabase project. A policy that looks correct but behaves differently at runtime is out of scope.
• Not an auto-fixer. The deterministic finding is the source of truth. The optional --ai fix is clearly labeled AI-drafted and must be reviewed before you apply it.

A clean rlsgate run means "none of the known high-signal holes are present," not "this app is secure."

License

MIT © 2026 Penguin Alley