Pkgtruth

pkgtruth

Ground truth about npm and PyPI packages, for AI coding agents and CI.

Your agent just wrote npm install unused-imports. That package is not the linter plugin it meant. It is a name an attacker registered because models kept inventing it — and npm has since replaced it with a security placeholder.

pkgtruth catches that before it reaches your lockfile — as an MCP server the agent asks, a CLI for CI, a GitHub Action on pull requests, or a Claude Code hook that denies the install command itself.

Why this exists

Large language models invent package names. Measured across models, 19.7% of generated package names were hallucinated, and when researchers re-ran the prompts, 43% of those names came back every single time.

That reproducibility is the whole attack. An attacker does not need to compromise a maintainer, poison a build server, or find a vulnerability. They watch what models invent, register the name, and wait. The technique is called slopsquatting, and it is already happening in the wild.

The standing security advice is that agents with package-management capabilities should not install anything without a review gate. pkgtruth is that gate, in a form an agent can call on its own.

Found in the wild

Two names a model plausibly produces, both live on npm today:

types-node is what you get when a model drops the scope from @types/node. npm removed it for malicious code in December 2024 and it is still installed ten times a week.

socket-io is not malicious — it is an abandoned package with a confusable name. That it takes 1,486 installs a week anyway is the point: a dot and a hyphen are enough.

PyPI has the same shape without the placeholder:

Neither is spelled anything like the package it stands in for, which is why pkgtruth also reads what a deprecation notice *says*: when it names a far more popular package, that name is the evidence.

npx pkgtruth check types-node socket-io

npx pkgtruth check -e pypi sklearn pytorch

Those two came from a hand check of twenty names. The systematic version — 242 well-known packages, every plausible garbling of each, regenerated every Monday — lives in [SLOPSQUATS.md](SLOPSQUATS.md). Section A there is npm security placeholders alone: names npm purged for malware that are still being installed this week. Weeks in which names enter or leave the list are published as releases, so the releases feed is a way to follow the live list without watching the repository.

Install

As an MCP server (for coding agents)

{

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

Three tools become available:

As a Claude Code hook (the install command itself is stopped)

An MCP tool only helps when the agent remembers to call it. A PreToolUse hook runs on every shell command instead, and denies npm install, npx, pnpm add, yarn add, bun add, pip install, uv add, poetry add and friends when a package they name is hallucinated or dangerous. Add this to ~/.claude/settings.json (every project) or .claude/settings.json (one repository, committable):

{

"hooks": { "PreToolUse": [ { "matcher": "Bash", "hooks": [{ "type": "command", "command": "npx -y pkgtruth hook" }] } ] } }

Commands that install nothing by name — a bare npm install, git, tests, npx of a tool already in node_modules — pass through with no network call. A blocked command comes back to the agent as a denial with the evidence, so it can pick the real package instead:

pkgtruth blocked this command: 1 of 2 package(s) must not be installed as-is.

✖ crossenv DANGER — npm replaced this package with a security placeholder (0.0.2-security). The name was used to publish malicious code. Package is 1 edit(s) from "cross-env", which has 22,155,059 weekly downloads (15,418x this one). Confirm you meant this package and not that one.

Add --fail-on caution to also stop packages that could not be verified. npm install -g pkgtruth with "command": "pkgtruth hook" skips the npx resolution on every call (measured: ~0.2 s per command via npx, ~0.03 s direct). The hook reads tool_input.command from Claude Code's JSON, or a bare command string from any other agent that pipes one in, and exits 2 with a permissionDecision: "deny" when it blocks — there is no environment variable that turns it off, because the agent controls the environment of the command it runs.

As a Claude Code plugin (server + hook in one install)

/plugin marketplace add hxckya/pkgtruth

/plugin install pkgtruth@pkgtruth

The plugin (plugin/) registers the MCP server and the hook above together, pinned to the npm release it ships with.

As a CLI (for humans and CI)

npx pkgtruth check express unused-imports          # npm (default)

npx pkgtruth check -e pypi requests sklearn # PyPI npx pkgtruth scan . # every manifest in the directory

scan reads package.json, requirements*.txt and pyproject.toml (PEP 621 and Poetry), checks each against its own registry, and exits non-zero when something is blocking, so it drops straight into CI.

As a pull-request gate (GitHub Action)

hxckya/pkgtruth-action (on the GitHub Marketplace) runs the scan on every pull request, posts one sticky comment with the evidence, and fails the check on HALLUCINATED or DANGER:

on:

pull_request: paths: ['package.json', '**/package.json'] permissions: contents: read pull-requests: write jobs: pkgtruth: runs-on: ubuntu-latest steps:

• uses: actions/checkout@v5
• uses: hxckya/pkgtruth-action@v1

This repository gates itself with it — see .github/workflows/gate.yml, which also proves the gate can fail by running it against a deliberately bad fixture.

What it checks

Verdicts are SAFE, CAUTION, DANGER, HALLUCINATED, or UNKNOWN. Every one arrives with the evidence behind it — an agent should never have to take "DANGER" on faith, and neither should you.

Design notes

Network failures never open the gate. If the registry is unreachable, the verdict is UNKNOWN, never SAFE. A degraded network must not silently turn a security check into a no-op.

Popular packages are not flagged — measured, not asserted. The false-positive audit runs the detector over the direct dependency closure of 240 well-known packages: 1,523 real packages, 0 blocking verdicts, 0 UNKNOWN after the built-in second pass, and 40 CAUTION (deprecated or unmaintained — true statements, not blocked by default). A gate that cries wolf gets switched off; this one has a number attached.

No build step. Two direct dependencies — the MCP SDK and zod, both only needed for the server. npx pkgtruth starts immediately.

Limitations

Read these before trusting it:

• npm and PyPI only. crates.io, Go modules, RubyGems are not covered yet.
• PyPI has no purge marker. npm leaves a -security placeholder where it removed malware; PyPI deletes the project, so a purged PyPI name simply reads as HALLUCINATED. PyPI also has no search API — near-twins are found against a daily snapshot of the top 15,000 projects by downloads, so an impostor of an obscure package will not be caught.
• Registry metadata only. It does not analyze package source code, so a legitimate-looking package with a malicious payload can still pass.
• Not a replacement for `npm audit` or Snyk. Those find known CVEs in code you already trust. pkgtruth asks the earlier question: should this package be here at all?
• New legitimate packages will get `CAUTION`. That is deliberate. Newness genuinely is a risk signal; use --fail-on danger so it does not block.

Options

--json              Machine-readable output

--fail-on danger (default) | caution

--fail-on caution also blocks packages that could not be verified at all, since "we could not check" is not a pass.

Exit codes: 0 clean, 1 blocking packages found, 2 usage or runtime error.

Configuration

On speed and rate limits

Adoption figures come from npm's downloads API, which throttles bursts and cannot batch scoped names — a project with several @scope/pkg dependencies would spend its whole budget on every scan.

Three things keep that in check: the bulk endpoint resolves all unscoped names in one request, requests to that host are paced serially, and figures are cached on disk for six hours. Weekly download counts move slowly, so a six-hour-old number is no less true.

A warm scan of ~18 dependencies takes about 1.4 second