vetpkg

Stop AI coding agents installing packages that don't exist.

Zero dependencies. Works as an MCP server, a CLI, or a library.

npx vetpkg express-async-router-utils-pro

 none  express-async-router-utils-pro         risk 100
        ! No package by this name exists on the registry. If an AI assistant
          suggested it, it was invented.
        ! Closest real package is "express-async-router" — every word in the
          shorter name appears in "express-async-router".
        => Do not install. The intended package was most likely "express-async-router".

The problem

AI coding assistants recommend packages that do not exist about 19.7% of the time (576,000-sample study). Open-weight models average 21.7%; commercial models around 5.2%.

That alone would just be annoying. What makes it dangerous is that the hallucinations repeat. When researchers re-ran 500 prompts that had produced a fabricated package name, 43% returned the same fabricated name every single time.

A predictable hallucination is an attackable one. The attack is called slopsquatting:

1. A model invents a plausible package name that has never existed.

2. An attacker registers that exact name on npm.

3. The next developer — or autonomous agent — who runs the install command pulls attacker-controlled code straight into the build.

This is not hypothetical. unused-imports is the textbook case: models hallucinate it instead of the real eslint-plugin-unused-imports. react-codeshift was claimed in January 2026.

Meanwhile over 50% of organisations have already had an outage or security incident from AI-generated code using outdated or wrong APIs (Snyk). Models are trained infrequently — a model shipped this summer can be confidently wrong about a library that changed a year ago.

Related MCP server: npmguard

What vetpkg does

It answers one question, before anything gets installed:

Is this a real, established package — or did a model invent it?

It checks the live registry and scores what it finds:

No single signal is proof. The giveaway is the combination — brand new, barely downloaded, published once, no repository, and named a near-miss of something popular. Real packages essentially never look like that.

Why name similarity needs three mechanisms

Edit distance alone misses the most common real-world shape:

The subset shape is exactly what models produce when they "simplify" a name they half-remember, and it's the shape of the documented unused-imports incident. A tool checking only edit distance will miss it.

Use it as an MCP server

This is the point. A CLI only helps if you remember to run it. An MCP server sits inside the agent's tool loop, so the check happens automatically, at the moment the agent is about to install something.

Claude Desktop / Claude Code — add to claude_desktop_config.json:

{
  "mcpServers": {
    "vetpkg": {
      "command": "npx",
      "args": ["-y", "vetpkg", "vetpkg-mcp"]
    }
  }
}

Cursor — add to .cursor/mcp.json:

{
  "mcpServers": {
    "vetpkg": { "command": "npx", "args": ["-y", "vetpkg", "vetpkg-mcp"] }
  }
}

Two tools are exposed:

• check_package — verify one package

• check_packages — verify several at once

Then tell your agent, once: "Always check packages with vetpkg before installing them."

Use it as a CLI

# check specific packages
npx vetpkg express react lodash

# check everything in package.json
npx vetpkg scan

# production dependencies only
npx vetpkg scan --prod

# machine-readable
npx vetpkg express --json

Exits non-zero if anything is suspicious, blocked, or missing — so it works as a CI gate or a pre-install hook:

{ "scripts": { "preinstall": "vetpkg scan || true" } }

Use it as a library

import { checkPackage } from 'vetpkg';

const result = await checkPackage('unused-imports');

result.verdict;         // 'trusted' | 'caution' | 'suspicious' | 'blocked' | 'not_found'
result.risk;            // 0–100
result.reasons;         // [{ severity, text, points }]
result.recommendation;  // plain-English next step
result.similar;         // closest real packages, if any

FAQ

Does this replace npm audit or Socket? No, and it isn't trying to. npm audit finds known vulnerabilities in packages you already trust. vetpkg asks whether the package should be trusted at all — whether it's real, and whether it's what you meant. Run both.

Why not just check if the package exists? Because the dangerous case is that it does exist. Once an attacker registers a hallucinated name, an existence check passes and hands you the payload. Existence is the easy half.

Will it flag legitimate new packages? Sometimes, and deliberately. A brand-new package with no downloads is genuinely unverifiable — that's caution, not suspicious. The scoring only reaches suspicious when several independent signals agree.

Does it flag popular packages with similar names? No. Packages with real adoption are exempt from name-similarity scoring — express is not warned about for resembling cypress. Slopsquats are by construction new and unused, so similarity only carries risk for packages that aren't established.

Does it send my dependency list anywhere? No. It queries the public npm registry directly from your machine. There is no telemetry and no account.

Which registries are supported? npm today. PyPI is next — the same attack works there, and the research covering it sampled both.

Why zero dependencies? A tool that vets your dependency tree shouldn't arrive with one.

How scoring works

Limitations

Worth being straight about:

• A determined attacker can fake the signals. Downloads can be inflated and repositories can be fabricated. This raises the cost of the attack; it does not make it impossible.

• The bundled name corpus is a snapshot. A squat targeting a package that isn't in the corpus won't be caught by similarity — though the age, adoption and provenance signals still apply.

• New legitimate packages score low. That's correct behaviour, not a bug, but it means low scores need judgement rather than blind obedience.

Licence

MIT