"Ok to proceed? (y/n)"is not security.

NPXGuard replaces npm's blind install prompt with a real risk report — publisher history, install scripts, obfuscation flags, and an AI-audited safety score. Before the code runs. Free CLI, one command.

Get the API →

Gateway evidence

23/100 · DANGER · BLOCKED

Blocked. Nothing executed.

npxguard · some-pdf-tool@2.1.4

$ npxguard some-pdf-tool

✗ CRITICAL postinstall script added in this version

✗ CRITICAL publisher changed 3 days ago (was: original-author)

✗ WARN network + env access pattern detected

✓ 8 other signals clean

You're one typo away from running a stranger's code as yourself.

  • Every npx command in every skill.md, README, and CI script is unaudited remote code execution. You're piping the internet into a shell with your credentials.
  • npm shows you a version number and asks yes/no. That's the entire security review. A package published 3 days ago by an account that took over last week gets the same prompt as lodash.
  • Your AI agents run these commands hundreds of times a day with nobody watching. They don't get suspicious. They don't hesitate. They type y.

The difference, verbatim.

What npm gives you

$ npx some-pdf-tool
Need to install the following packages:
  some-pdf-tool@2.1.4
Ok to proceed? (y)

That's it. That's the whole security model.

What NPXGuard gives you

$ npxguard npx some-pdf-tool

NPXGuard: some-pdf-tool@2.1.4

SCORE  23/100   VERDICT: DANGER

  x publisherChanged   version published 3 days ago by an account
                       that isn't the previous publisher
  x installScripts     postinstall script ADDED in this version
  x envExfil           reads process.env in a file that also opens
                       a network connection — exfiltration pattern
  x downloadsUnder1k   214 weekly downloads — little vetting

  Full report: npxguard.dev/p/some-pdf-tool/2.1.4

Blocked. Run with --force if you enjoy incident reports.

Nothing executed. You just found out before your env vars did.

How it works

  1. 1. Wrap.

    npxguard <anything> — wrap npx, wrap npm install, wrap the sketchy command a README told you to run. Zero config.

  2. 2. Analyze.

    12 static signals — publisher history, install scripts, obfuscation, typosquat distance, network and filesystem access, download patterns — in under 2 seconds. The package is parsed, never run. Nothing executes until you approve.

  3. 3. Decide.

    You (or your agent) get the score and the receipts: every flag, every penalty, every file that triggered it. Not a vibe. A report.

Try it now

Score any npm package right here — no account, no install. This is the same anonymous endpoint the free CLI uses.

Your agents install packages. Who's checking them?

Your coding agent will npm install anything that looks plausible. It has your tokens, your SSH keys, and no survival instinct.

One API call before every install.

POST /check returns allowed: true/false with a machine-readable reason — closed enums, structured violations, no prose your agent has to squint at.

Policy threshold.

Set your line (block on DANGER, warn on CAUTION). Below it, the agent stops and asks a human. Every decision is audit-logged.

Five lines into any runtime.

The SDK drops into Claude Code skills, LangChain tools, or a bare fetch call. If NPXGuard is unreachable, the answer is allowed: false — the gate fails closed, never open.

Durable scans.

Kick off a deep AI analysis, poll the job, never block. Your agent keeps working; the verdict lands when it lands.

The CLI sends one thing: the package name.

Never your code. Never your files. Never your env.

That's not a settings toggle, it's the architecture — the CLI has no code path that reads your project files. We analyze the package from the registry, not from your machine. Crash reports (opt-out) contain a stack trace and nothing else.

Pricing

Free means free — the CLI is the product's front door, not a trial.

Free

$0

every developer

  • 100 standard scans per day
  • Up to approximately 3,000 scans per month, with daily limits.
  • CLI
  • Shareable reports
  • Community support
Install the CLI

Pro

$29/mo

power users & CI

  • 5,000 scans/mo API
  • 100 deep AI analyses/mo
  • 10 req/s
  • Agent /check gate
  • Email support
Start with Pro

Team

$99/mo

teams shipping daily

  • 50,000 scans/mo API
  • 300 deep AI analyses/mo
  • 20 req/s
  • Agent /check gate
  • Priority support
Opening soon

Agent

$299/mo

autonomous fleets

  • Metered, 50 req/s
  • 1,000 deep AI analyses/mo
  • Policy engine + audit log
  • Priority + Slack support
Launch access

FAQ

How is this different from Socket?

Socket is a great dependency-review platform for your repos. NPXGuard guards the moment of execution — the npx command you're about to run right now, and the installs your agents make autonomously. Pre-flight gate, not post-hoc dashboard. Use both; we won't be jealous.

Does it slow my installs down?

Cache hit: under a second — popular packages are already scored globally. Cache miss: about 2 seconds of static analysis. Deep AI analysis runs as a background job and never blocks an install. You've waited longer for npm to print its funding banner.

Does it work with private registries?

Not at launch. NPXGuard analyzes packages from the public npm registry. Private registry and scoped-proxy support is on the roadmap — tell us what you run and we'll prioritize.

Can it block installs in CI?

Yes. The CLI exits nonzero on DANGER (distinct exit codes for CAUTION and service-unreachable, so you choose the policy). One line in your pipeline: npxguard check $PKG before the install step.

What data do you collect?

Package name and version — that's what a scan needs. With an account: your scan history and API usage (that's your quota). Opt-out crash reports. We never see your code, your files, or your environment.

Do you execute the packages you analyze?

Never. Tarballs are parsed in memory as data and discarded — no install scripts run, no code is evaluated, ever. It is the first invariant in our security architecture and the build fails if anyone tries to change it.