Checks a Vercel + Supabase repo before deploy: RLS, grants, leaked keys, webhooks. Runs locally.
io.github.ohad6k/viberaven — VibeRaven MCP Server
VibeRaven MCP (io.github.ohad6k/viberaven) is an MCP server focused on production-readiness checks and Launch Autopilot next actions. The server’s scope is defined by its description: it provides functionality to evaluate whether production is ready and to determine subsequent actions for Launch Autopilot.
🛠️ Key Features
Production-readiness checks
Launch Autopilot next actions
🚀 Use Cases
Verifying production readiness before deployment
Determining what Launch Autopilot should do next
⚡ Developer Benefits
Exposes production-readiness and next-action logic via an MCP server interface
Uses a focused workflow centered on readiness checks and Launch Autopilot sequencing
⚠️ Limitations
Available metadata does not specify supported tools, endpoints, protocols, or detailed behavior beyond the described responsibilities.
The production layer for AI-built apps.
Checks across Supabase, Vercel and Stripe, and an engine that races every rule in your database on real PostgreSQL, before your users do.
The engine inside VibeRaven, on the bundled example:
sh
npx -y viberaven@1.10.0 race --example
text
At most 2 bookings per event: 60 orders, 24 broken, 0 waiting; 1.495s total
read committed: 12/20 broken
repeatable read: 12/20 broken
serializable: 0/20 broken
First broken order 001101 (read committed); COMMIT A=true, B=true
Replay record is byte-identical, including PostgreSQL protocol bytes.
That is the engine inside VibeRaven: race tests your app's database with two users at once, on real PostgreSQL, without Docker, and replays exactly what broke. In order 001101, A inserts, B inserts, A commits, B commits: the event ends with 3 bookings out of 2. The bundled fix brings it to 0 broken of 60.
npx -y viberaven race --example: two users book the last seat at once. 12 orders broke under read committed, 12 under repeatable read, 0 under serializable: 24 of 60. The broken order replays exactly.
Inside your chat
Add https://viberaven.dev/mcp as a custom connector in ChatGPT, Claude, Codex, Cursor or Lovable, then ask: "Can two users book the last seat?" Setup for each agent.
In your repo
sh
npx -y viberaven@1.10.0 check # what is risky before launch
npx -y viberaven@1.10.0 race # what breaks when two users act at once
check reads migrations, client code and env files, and names the file and line for each gap.
race runs the rule in your viberaven.race.json: setup, two requests and the SQL that must stay true. Spec template.
Fix, run race again, and replay the reported order. Exit 0: no order broke. Exit 1: broken. Exit 2: incomplete.
race needs Node 24+ (the CLI enables JSPI). check runs on Node 20+.
Add https://viberaven.dev/mcp as a custom connector
Local MCP tools: npx -y @viberaven/mcp@1.10.0. Details in the agent hub.
Building an agent that answers provider errors? The AI SDK provider-context example adds search_guides and recent_provider_changes from the hosted MCP endpoint.
Local Studio
npx -y viberaven@1.10.0 opens the local Studio: chat with your coding agent, the provider board, versions and diffs, and a Race test tab that runs race and replays a broken order.
check reads the repository. It does not show which policies are live in Supabase or Vercel, and a clear result is not a security audit.
race covers the SQL, inputs and bounded statement orders you supply. Rules beyond a simple count limit need an explicit spec. Omitted roles run as postgres.
An incomplete run is not a pass. Replay needs the same migrations, spec, options and runtime.
Cloud Auth, HTTP and external I/O are outside the test. Run only SQL you trust; the engine is not a security sandbox.
No universal PostgreSQL compatibility, and no advantage over a plain coding agent in bug finding, is claimed.