In private testing · 12 agents · Postgres, MySQL, SQLite

Vibecode at full speed without getting wrecked.

Rollcage snapshots your production database before an AI agent can break it, then asks you first. Everything else goes straight through in about 4 ms.

Or check a command of your own: Rollcage's real engine runs right here, in your browser.

Claude CodeCursorCodexCopilotDevinAntigravityGemini CLIClineopencodeAmpDroidQwen Code SupabaseNeonRailwayAWS RDSMySQLMariaDBSQLitePrismaRailsDjangoMCP servers

Why

It's already happened.

AI agents move fast, sound sure of themselves, and have already deleted real production databases.

July 2025 Code freeze. Database gone.

Replit's agent was told not to change anything. It ran destructive commands against production anyway.

April 2026 One call. Backups too.

A Cursor agent deleted PocketOS's production database on Railway, along with the backups kept on the same volume.

With Rollcage The way back stays out of reach.

A snapshot before every risky command, encrypted with your key, stored where you choose, never next to the database it protects.

Simulator

Go on. Try to break production.

A real-looking agent, a real-looking table, and a bad idea of your choosing. Run it with Rollcage off, then on.

~/acme-app · claude
The bad idea
Agent
The users table is a mess. Clean up the database so we can start fresh.
users production · Supabase
48,213 rows
idemailplanjoined
ERROR: relation "users" does not exist

48,213 customers, gone. The backups were on the same volume.

Rollcage is on. Run the agent and watch what happens.

How it works

From risky command to safe, in one beat.

agent psql $DATABASE_URL -c "DROP TABLE users"
agent npx prisma migrate reset --force
agent supabase db reset --linked
agent railway volume delete pg-data
agent DELETE FROM orders;
ls -lasafe3.9 ms
npm testsafe4.1 ms
psql … "DROP TABLE users"production4.3 ms
git statussafe4.0 ms
pg_dumpproduction
zstdcompressed
ageencrypted
Your bucketrc_20261117_0412
Rollcage

This would DROP TABLE users on production (Supabase).

Snapshot saved: rc_20261117_0412. Allow it?

$ rollcage restore rc_20261117_0412
Creating app_restored_20261117_0412...
Downloading and decrypting...
Restoring...
Restored. Your original database wasn't touched.

Command X-ray

See exactly what Rollcage sees.

Real verdicts from the Rollcage engine, recorded with rollcage explain for a project whose production database is on Supabase. Search them, or type any command: the same engine checks it live, right in your browser.

Recorded with Rollcage's own engine. Nothing was run.Edit it in the playground

The sample project

Every check here, recorded or live, runs in this project. Rollcage reads its files the way it would on your machine: which databases are protected, the connection strings in .env, and what the scripts in package.json and the Makefile really run. To change them (add a rule of your own, say), open the playground.


        

Agents

Whatever you code with.

Rollcage hooks into each agent the way that agent allows. Pick yours to see where the hook lives and what you'll see on a close call.

Asks you on a close call Blocks (or waits for a dashboard approval)

Features

Real guardrails, not vibes.

Rollcage reads the command the way your shell would, then works out what it would actually touch.

Finds SQL wherever it hides.

Inside psql -c, SQL files, heredocs, npm run scripts, bash -c, eval, Python and Node one-liners, Prisma and Drizzle, and MCP tools. Plus the framework commands that wipe a database, and the migrations an agent just wrote.

Knows which database is production.

By connection string, host, project ref or env var. Your local dev database stays out of the way.

Your snapshots, your storage.

Encrypted on your machine with a key only you hold. Kept in R2, S3, Spaces or a local folder.

Restores that don't make it worse.

Into a new database by default, or just the one table you lost. rollcage undo finds what a bad command took and brings it back.

A log of every close call.

What was allowed, asked about or blocked, with the snapshot it took. Secrets redacted.

Built for the worst day.

Rollcage assumes the agent is fast, confident and occasionally wrong, so it closes the gaps a clever agent might find.

  • No snapshot, no commandIf the safety snapshot fails, the command is blocked. Nothing runs unprotected.
  • Agents can't touch the safety netDeleting snapshots, reading your key or editing Rollcage's settings always asks you first.
  • Fails closedIf Rollcage can't check a risky-looking command, it blocks it. It never gets in the way of safe ones.
  • Secrets stay secretPasswords, keys and tokens are redacted from every message and log line.
  • Bypass mode tooPermissions skipped? Rollcage still snapshots, then blocks instead of asking.
  • Your rules, tooAdd your own patterns in rollcage.toml: warn, ask or block outright.

Why not just...

You've probably thought of these.

Each one helps. None of them stops the one command that matters, at the moment it runs.

...use the agent's own permission prompts?

The catchThey ask about everything, so before long most people switch on auto-approve or skip permissions. And a prompt shows a command, not what it touches: npm run db:reset doesn't say production.

RollcageStays quiet for safe commands, reads through scripts and .env files to say exactly which database a command would hit, and still snapshots with permissions skipped.

...keep backups?

The catchLast night's backup loses today's work, restoring it rolls back everything else too, and backups kept next to the database go down with it. PocketOS's did.

RollcageA snapshot from the second before, encrypted with your key, kept where you choose. Bring back the one table you lost, into a new database.

...give the agent a read-only database user?

The catchThen it can't run the migration, the seed or the fix you asked for. And it reads the same .env you do, so it usually has your full-rights DATABASE_URL anyway.

RollcageEverything runs, except the destructive few. Those wait for a snapshot and your yes.

...tell the agent not to?

The catchInstructions are suggestions to a model. Replit's agent was told about a code freeze and ran destructive commands against production anyway.

RollcageA hook every command goes through, outside the model. It can't be talked out of it, and agents can't edit its settings without asking you.

Speed

About 4 ms. You'll never notice it.

Safe commands pass straight through without touching the network. Rollcage only slows down when it's saving your data.

4.1 ms

A typical check on an Apple Silicon Mac, from start to exit. The slowest 5% take about 10.5 ms; the budget is 20.

Release build, whole process, CI benchmarkbudget 20 ms

Config builder

Your rollcage.toml, before you install.

rollcage init writes this for you from a settings page. Here's what it'll look like for your setup. It's safe to commit: no passwords, keys or connection strings.

Production database
Connection string lives in
Snapshots go to
Keep snapshots for
rollcage.toml
Try it against this configRollcage's real engine, in your browser

Pick a command, or type one: the engine checks it against the config above, with a .env holding the connection string. Change the config and the answer follows.

At launch

Set up once. Then forget it's there.

Rollcage is in private testing. When it opens, setup takes about two minutes.

  1. 01

    Install with one line

    Downloads Rollcage for your Mac or Linux machine and checks the download.

    curl -fsSL https://rollcage.dev/install.sh | sh
  2. 02

    Pick what to protect

    A settings page opens in your browser. Choose your production database, where snapshots go, and folders your agent should never delete.

    rollcage init
  3. 03

    Keep vibecoding

    Start your agent like always. Rollcage only speaks up when your data is on the line.

    claude

Optional dashboard

Every close call, in one place.

  • Approve from your phone. Agents that can't ask you wait for your answer on the dashboard.
  • Connect with a code. Your terminal shows it, you click Connect. No passwords typed anywhere.
  • Metadata only. Database names, what was caught, snapshot sizes. Never connection strings, keys or your data.

FAQ

Questions.

A small command-line tool that sits between your AI coding agent and your production database. It checks every command the agent runs, takes an encrypted snapshot before anything destructive reaches production, and asks you before it runs.

Rollcage is in private testing right now. Join the waitlist and you'll get one email when it opens. Installing will take one line, and a settings page walks you through the rest.

Claude Code, Cursor, Codex, GitHub Copilot (the CLI, and its agent in VS Code), Devin (formerly Windsurf), Antigravity (the app, its IDE and the agy CLI that replaced Gemini CLI), Gemini CLI (still around with paid API keys), Cline, opencode, Amp, Droid (Factory) and Qwen Code. Databases: Postgres (Supabase, Neon, Railway, RDS and any other Postgres you can reach with a connection string), MySQL or MariaDB, and SQLite.

No. Safe commands are checked in about 4 milliseconds without any network calls. The only pause is when a destructive command is aimed at production and Rollcage is saving a snapshot.

Wherever you choose: Cloudflare R2, AWS S3, DigitalOcean Spaces, any S3-compatible storage, or a folder on your computer. They're encrypted before they leave your machine, with a key that only you have.

Say yes. Rollcage isn't there to stop you working. It's there to make sure there's a way back. The snapshot is already saved, and every decision lands in the activity log.

No. Rollcage runs on your computer. Snapshots are encrypted on your machine and go straight to the storage you picked. If you connect the optional dashboard, it gets metadata only: database names and hosts, what Rollcage caught, snapshot ids and sizes, and table names with their row counts. Never connection strings, keys, or any row of your data.

Yes, right on this page. Type any command into the Command X-ray, or build your rollcage.toml in the config builder and try commands against it. Both run Rollcage's real engine, compiled to WebAssembly, in your browser: the same code the hook runs, so you get the answer Rollcage would give in a project set up like that. In the playground you can change the project itself: add a rule of your own, protect a folder, change what a script runs. Nothing you type is sent anywhere.

Run rollcage undo. It finds the snapshot Rollcage took just before that command, shows which tables were dropped or emptied since, and brings them back into a new database. Then rollcage copy-back --preview shows exactly what copying back would do to today's data, and rollcage copy-back puts the missing rows back into production once you confirm, keeping rows added since.

Not quietly. Changing Rollcage's settings or hooks, replacing the Rollcage program, reading its key, deleting snapshots, or starting another agent with its hooks off always asks you first, however it's attempted. Your own allow rules can't switch that off.

Run rollcage explain '<the command>'. It shows what Rollcage would do with it and why: which rule matched, which database it's aimed at, and whether a snapshot comes first. Nothing is run. The Command X-ray above is made from exactly that.

Go fast. Keep your data.

Rollcage is in private testing. Join the waitlist and we'll email you when it's ready for your project.

One email when Rollcage opens. We never share your address.