Bumper
Beta

Your AI is about to do something
risky

Bumper pauses it, explains what's about to happen in plain English, and waits for you to say yes or no.

npm install -g bumper-guard
Free — early access

Sign in with just your email, no password.

See it in action

The same command. Two different outcomes.

Without Bumper runs blindly
$rm -rf ./public/images
Deleting…
Done.
0 questions asked images folder gone for good
Bumper catches it first
Why this exists

Vibe-coded apps are shipping with the front door left open.

When AI writes most of the code, nobody's necessarily reading every line before it runs — least of all the risky lines.

5,600
publicly deployed vibe-coded apps scanned by security researchers
400+
had real, live API keys or secrets exposed in the shipped code
2,000+
critical vulnerabilities found across the same set of apps

Source: Escape.tech's early-2026 scan of publicly deployed vibe-coded applications.

How it works

Three steps. No new habits to learn.

01

Install it once

One command wires Bumper into whichever AI coding tool you already use — no new app to open, no workflow to change.

02

It watches from inside your agent

Bumper plugs into your agent's own permission system — the same mechanism that already asks "run this command?" — not a separate app watching from the outside.

03

Risky stuff pauses for you

Safe things run instantly. Genuinely risky things stop and wait for a plain-English yes or no from you, right when it matters.

What it actually checks

The default rules, in plain terms.

Everything below runs locally out of the box. You can add your own rules — these are just the ones most vibe-coded apps get burned by.

Blocked automatically

Deleting everything, no undo

rm -rf *
Asks first

Rewriting shared history

git push --force
Asks first

Wiping data from a database

DROP TABLE ...
Asks first

A real secret key landing in your code

const key = "sk_live_..."
Where it works

One tool. Every agent you already switch between.

Most safety tools protect one agent, because they'd have to be built by the same company that makes it. Bumper is wired into all three separately, through each one's own native permission system — not a bolt-on watcher hoping to catch what leaks through.

Claude Code

The first integration built — verified end to end before any other agent was added.

Cursor

Verified against a real, signed-in Cursor account — not just Cursor's published hook schema.

Copilot CLI

Verified against a real, GitHub-authenticated Copilot session, not a simulated one.

Pricing

Free during early access.

Sign in with your email — a code, no password to remember — and everything's unlimited while this is early. Real pricing comes later, set from how people actually use Bumper instead of a guess.

Free — early access
$0
Everything unlocked, no catch
  • Every agent, every rule
  • Unlimited ask-decisions
  • Sign in with email — no password

No credit card, no plan to pick — run bumper login and you're in.

Install & guide

Everything worth knowing before you turn this on.

Install

Same two commands regardless of which agent you use. Run the install once per project (or add --global to cover every project on the machine).

1
$npm install -g bumper-guard
$bumper start
Starts the local daemon — leave it running in the background. Don't want to remember? Run bumper autostart enable once and it starts itself on login.
2
$bumper install claude-code
Using more than one agent? Run bumper install all to wire in every supported agent at once.
3
$bumper status
$bumper log
Check it's running, or see every decision it's made so far — automatic and yours.

How a decision actually gets made

Every risky-looking action lands in one of three buckets the instant your agent tries it.

Allow

Matches a rule marked safe — proceeds instantly, nothing shown to you.

Deny

Matches a rule marked unsafe — blocked automatically, your agent sees why.

Ask

Genuinely unclear — your agent pauses and shows Bumper's plain-English explanation right there in its own native permission prompt, waiting for your yes or no.

Advanced

Restrict something Bumper doesn't already catch.

For anyone who wants to go beyond the built-in danger pack — every project has its own risky command. Drop a bumper.policy.yaml in your project root (or ~/.bumper/policy.yaml for every project) and Bumper checks it first. Your rules add to the 50+ built-in ones — they never replace them.

rules:
- id: block-prod-migration
match:
command: "*prisma migrate deploy*"
decision: ask
explain: "This runs a database migration against production."

Match on command (glob), file (glob), or content_pattern (regex against code being written) — set decision to allow, deny, or ask, and either write explain yourself or reuse a built-in category (like recursive-delete or secret-write) to skip writing the sentence. Full reference: bumper.policy.example.yaml.

Privacy

Your code and commands never leave your machine.

An account is required — sign in with your email and a one-time code, no password to remember or leak. Once signed in, every check still runs locally against your own policy: your code and commands never leave the machine. Only a session token travels to Bumper to confirm you're logged in — never what you're actually running.

Protect the next thing you build.

One install, wired into the agent you already have open. Free during early access — sign in and you're protected in under a minute.

npm install -g bumper-guard