greenlitlab
# Greenlit
**Is your AI-built app safe to launch?** Greenlit grades a deployed app from A to F for what a stranger can use against it. It explains each problem in plain words and gives you a prompt to paste back into Lovable, Bolt, v0, Replit or Emergent that fixes it. Run it again after the fix, and it shows what got fixed.
<p><img src="https://raw.githubusercontent.com/priberry-data/greenlitlab/main/docs/demo.gif" alt="Greenlit checks a Lovable app: grade F with a leaked Stripe key, then a fix prompt to paste back" width="720"></p>
**In the browser:** [greenlit.priberry.com](https://greenlit.priberry.com). Paste your app's address. Nothing to install.
**On your computer:**
```bash
npx greenlitlab myapp.lovable.app
```
## What it checks
| Check | Where it runs |
|---|---|
| Secret keys in the JavaScript every visitor downloads: 32 kinds, including Stripe, OpenAI, Anthropic, OpenRouter, Gemini, AWS, GitHub, Supabase secret keys, Telegram bots, private keys and service-account JSON | everywhere |
| A Supabase `service_role` key in the browser | everywhere |
| Source maps that hand out your source code | everywhere |
| Clickjacking (no `frame-ancestors` or `X-Frame-Options`) and no Content-Security-Policy | everywhere |
| Supabase tables anyone can read with the public key | your own app: `--mine`, or on the website with an owner tag |
| Supabase storage buckets that list their files to anyone | your own app: `--mine`, or on the website with an owner tag |
| Supabase sign-up open to anyone without email confirmation | your own app: `--mine`, or on the website with an owner tag |
| A public `.env` file or `.git` folder | your own app: `--mine`, or on the website with an owner tag |
| Firebase Realtime Database and Firestore open to everyone | your own app: `--mine`, or on the website with an owner tag |
Greenlit reads all of the app's own scripts, including lazy-loaded pages and the app's own CDN, and tells you if a check ran out of room or couldn't read the code. Grades follow the danger: one critical problem is an F whatever else is right, one high problem means C at best, and small things such as missing headers take 20 points at most. What your host controls and you can't change is shown but costs nothing.
On the website, choose "Check my database here" after a check: you add a tag made from a key only your browser holds, so nobody else can use your tag. On the website it checks standard Supabase and Firebase addresses only.
Supabase no longer lists a project's tables for the public key, so the database check tests the tables your app's code names, tells you how many it tested, and says so plainly when it could test none.
The owner-only checks ask for a row count (Supabase), one file or document name (storage, Firebase), or the sign-up settings, never for data, and never write anything. They run only when you say the app is yours, because testing someone else's database without permission isn't OK.
## Launch readiness, beyond leaks
| Check | Where it runs |
|---|---|
| A page that loads from a developer-only address, plain-http content, Stripe in test mode, a default title, "noindex", AI calls from the browser, a test email sender, unlocked sign-in cookies | everywhere |
| Plain http that doesn't redirect to https; an app that lets any website read a signed-in visitor's data | everywhere |
| No DMARC record or a broken SPF record on your own domain | everywhere, custom domains only |
| Personal columns (email, phone, password…) in tables strangers can read; backend functions listed for review | your own app: `--mine`, or on the website with an owner tag |
| Rules that are too loose, tables strangers can change: Supabase's Security Advisor in plain words | `advisor`, with your advisor results |
| Always-true policies, tables without Row Level Security, unverified payment webhooks, secrets behind a public prefix | `scan`, on your project's files |
| Backups, spending limits, a separate test database, outage alerts | `questions`: only you know, never graded |
Only what Greenlit confirmed moves the grade. What it couldn't see is said out loud, and what only you know is asked. Greenlit never signs in, never reads your users' data and never writes to your database.
## Commands
```bash
npx greenlitlab <url> # grade an app
npx greenlitlab <url> --mine # your own app: database checks too
npx greenlitlab <url> --prompt 1 # print fix prompt 1 only (pipe it to pbcopy)
npx greenlitlab <url> --prompt all # every fix in one prompt, worst first
npx greenlitlab <url> --json # the full result
npx greenlitlab <url> --fail-on high # exit 1 if anything this serious is found, for CI
npx greenlitlab questions # seven things only you know, with a fix for each gap
npx greenlitlab scan [folder] # read your project's files before you deploy
npx greenlitlab advisor results.json # Supabase's Security Advisor, in plain words
```
The last result for each app is kept in `~/.greenlit/sites/` (grade, finding ids and titles only, never keys) so the next run can show what changed.
## In Claude, Cursor, VS Code and other agents
Add Greenlit as an MCP server:
```json
{ "mcpServers": { "greenlit": { "command": "npx", "args": ["-y", "greenlitlab@0.2.0", "mcp"] } } }
```
Then ask your agent to check the app it just deployed and apply the fixes. The `greenlit_check` tool returns the grade, every finding and its fix prompt.
## Made by Priberry
Greenlit comes from [Priberry](https://priberry.com), the crash-test lab for AI agents. Its sister tools: [Crumple](https://crumple.priberry.com) crash-tests coding agents, and [Tare](https://tare.priberry.com) shows what they cost.
Apache-2.0. Found a problem? [Open an issue](https://github.com/priberry-data/greenlitlab/issues), or see [SECURITY.md](SECURITY.md).
TDQS
Scored across 4 tools
Each tool has a clearly distinct input source and scope: local project files (greenlit_scan), Supabase advisor lints passed in (greenlit_advisor), owner-only answers (greenlit_questions), and a live deployed URL (greenlit_check). The descriptions explicitly delimit boundaries, e.g. scan covers what 'can't be seen from outside' while check covers 'what the site serves to every visitor', so an agent can pick correctly.
All four names share a uniform greenlit_ prefix with a single short lowercase suffix, giving a predictable pattern. Minor deviation: suffixes mix verbs (scan, check) and nouns (advisor, questions), so it isn't a strict verb_noun convention, but it remains consistent and readable.
Four tools for a focused launch-readiness audit product is well-scoped, with no redundant entries. It sits at the lean end of the 3-15 range, but each tool earns its place and none could be merged without losing a distinct input channel.
The surface covers the main launch-check layers: local code, external deployed site, database advisor lints, and owner-attested operational questions, each returning findings with fix prompts. Minor gaps exist, such as no consolidated report tool that aggregates findings across all four checks or persists results between runs beyond the per-run change diff.