MCP 服务器

BeforeShip MCP Server

Provides pre-deploy AI-assisted security and readiness checks for AI-built apps, including scanning for secrets, auth gaps, and payment safety.

README

BeforeShip

Your AI coding agent says the app is done. BeforeShip checks if it is safe to deploy.

BeforeShip is an open-source launch gate for solo founders and vibe coders building apps with Cursor, Claude Code, Codex, Replit, Lovable, Bolt, v0, and other AI coding tools.

It scans AI-built SaaS apps for the production basics coding agents often skip:

auth and authorization gaps
exposed secrets and unsafe env usage
Supabase RLS/database isolation issues
Stripe/Razorpay webhook safety
rate limits, CORS, validation, and security headers
missing smoke tests and deployment readiness
risky AI-agent diffs like deleted tests, huge rewrites, or auth/payment changes without verification

The goal is simple:

Run one command before deploy. Know if your vibe-coded app is safe to ship.

Status

This repository is starting as a public product spec and research-backed roadmap for an OpenAI Hackathon project.

The first executable slice is now in progress:

npm test
npm run scan -- --no-report
node packages/cli/bin/beforeship.js scan /path/to/app
node packages/cli/bin/beforeship.js loop /path/to/app --max-iterations 3

Current checks:

basic stack detection for Next.js, Supabase, Stripe/Razorpay, Vercel, and env files
committed env-file warning
secret-like token detection for OpenAI, Supabase JWT-style keys, Stripe, and Razorpay
Next.js API route auth-guard and trusted-user-id heuristics
Stripe/Razorpay webhook signature and idempotency heuristics
Supabase service-role, RLS, owner-policy, and public-bucket heuristics
API safety checks for wildcard CORS, auth-route rate limits, and input validation
Next.js deploy-readiness checks for tests, health routes, and .env.example
score/verdict engine
markdown and JSON reports under .beforeship/
bounded verification loop with named terminal states and loop memory

Demo fixture:

node packages/cli/bin/beforeship.js scan examples/flawed-next-supabase-app --no-report

The fixture should return DO NOT SHIP and show the launch blockers BeforeShip is designed to catch.

Repository-level scans respect .beforeshipignore, which excludes this intentionally flawed fixture.

The intended MVP is:

CLI first + MCP server + optional GitHub Action

Initial target stack:

Next.js + Supabase + Stripe/Razorpay + Vercel

Why BeforeShip?

AI coding tools make it easy to create working apps, but hard to verify production readiness.

A vibe-coded app can look 90% done while missing the dangerous invisible pieces:

payment webhook signature verification
database row-level security
server-side auth checks
safe API key handling
rate limiting
rollback and smoke tests
deployment env separation

BeforeShip is the missing pre-deploy gate between:

"It works locally"

and:

"It is safe to launch"

Product thesis

Coding agents optimize for making code run. BeforeShip checks whether the generated app is safe to ship.

BeforeShip is not meant to replace Semgrep, Snyk, CodeQL, Gitleaks, or human review.

It is meant to orchestrate and explain the checks solo founders actually need before deploying an AI-built SaaS app.

Planned CLI

npx beforeship scan
npx beforeship scan --url https://myapp.com
npx beforeship report
npx beforeship fix-prompts
npx beforeship mcp

Example output:

BeforeShip score: 42/100
Verdict: DO NOT SHIP

Critical:
1. Razorpay webhook does not verify signature
2. Supabase service role key appears in client code
3. /api/admin/users has no auth guard

High:
4. No rate limiting on login
5. CORS allows all origins
6. No smoke tests found

Planned MCP tools

BeforeShip should expose a small number of high-leverage MCP tools, not a huge overloaded tool list.

before_ship_scan
before_ship_explain
before_ship_fix_plan
before_ship_gate_deploy

Optional:

before_ship_agent_policy

Planned checks

Secrets

.env committed
API keys in frontend bundle
NEXT_PUBLIC_ misuse
Supabase service role exposed
OpenAI/Anthropic/GitHub keys exposed
Stripe/Razorpay secrets exposed
Cursor/agent context includes secret files

Auth

API routes without session/auth guard
admin pages without server-side protection
user ID trusted from body/query
IDOR patterns
auth middleware exists but is unused
role checks performed client-side only

Supabase / database

RLS disabled
tables with user data and no owner policy
service role used outside server
public anon key misuse
migration drift
unsafe storage buckets

Payments

Stripe/Razorpay webhook signature missing
missing idempotency
subscription status trusted from frontend
price/plan selected client-side
billing country/currency not server-validated
entitlement table missing
payment success page grants access without verified webhook

API safety

wildcard CORS
no rate limiting
no input validation
no CSRF for cookie auth flows
verbose errors
insecure file uploads
missing security headers
SSRF patterns

Deploy readiness

no health endpoint
no smoke tests
no rollback note
missing env vars
Vercel serverless timeout risks
critical routes lack logging
no staging/prod separation

AI-agent-specific checks

huge AI-generated files
repeated full-file rewrites
deleted tests
new dependency added without reason
changed auth/payment/db files without tests
destructive scripts/migrations
no AGENTS.md / CLAUDE.md
agent touched files outside declared scope
diff contains temporary auth bypasses
tests modified to fit broken code

Planned report files

.beforeship/
  config.yml
  launch-report.md
  findings.json
  fix-prompts.md
  evidence/

Example finding:

{
  "id": "payment.webhook.signature_missing",
  "severity": "critical",
  "file": "web/src/app/api/razorpay/webhook/route.ts",
  "title": "Razorpay webhook does not verify signature",
  "why_it_matters": "Anyone can fake a payment success event and unlock access.",
  "fix_prompt": "Add Razorpay webhook signature verification using RAZORPAY_WEBHOOK_SECRET. Reject unsigned or invalid payloads. Add a regression test."
}

How OpenAI fits

BeforeShip should use deterministic checks first and OpenAI reasoning second.

OpenAI can help with:

stack-aware risk triage
founder-friendly explanations
fix prompt generation
test plan generation
false-positive reduction
deploy-gate reasoning for AI agents

It should not rely only on LLM judgement for security.

Repository structure

Current repository is documentation-first.

Planned implementation shape:

beforeship/
  packages/
    cli/
    core/
    mcp/
    github-action/
  docs/
    research.md
    product-spec.md
    roadmap.md
  examples/
    flawed-next-supabase-app/