@guardbee/mcp-slopsquat-scanner
🇬🇧 English | 🇹🇷 Türkçe | 🇨🇳 中文 | 🇯🇵 日本語 | 🇰🇷 한국어 | 🇧🇷 Português | 🇪🇸 Español | 🇩🇪 Deutsch
[npm version]
[npm downloads]
[License: MIT]
An MCP (Model Context Protocol) server that checks every dependency declared in a project's own manifest (package.json, requirements.txt, pyproject.toml) against the real npm/PyPI registry — not a local database, an actual network lookup.
LLM coding assistants hallucinate plausible-sounding package names at a measurable rate — research on open-source models puts it around 21.7%. Attackers know this and pre-register exactly those names on npm/PyPI, so the next person who copies the same suggestion doesn't get an install error — they get the attacker's code instead. This is slopsquatting, and unlike a typo-squat, there's no misspelling to catch by eye: the hallucinated name is often more plausible-sounding than the real package it was meant to be.
Telemetry is off by default. With GUARDBEE_TELEMETRY=1 it sends the server, tool, success, duration and short parameter values such as a file path; longer strings and fields like content, token or password are replaced with a placeholder. See @guardbee/mcp-telemetry.
Why a live registry lookup, not a local list
There's no way to maintain a local "known bad" list for this — the whole point of a slopsquat is that it's a new name, registered specifically to catch a specific hallucination, possibly hours before someone runs npm install. A static pattern can't know that; only asking the registry "does this exact name exist right now" can. This package only checks a project's own directly declared dependencies (the dependencies/devDependencies/peerDependencies/optionalDependencies sections of package.json, or requirements.txt/pyproject.toml for Python) — not the full resolved lockfile tree. Transitive dependencies were already vetted by whichever maintainer published the package that depends on them; the actual risk moment is when a new name gets typed (by a human or an AI assistant) directly into your own manifest.
A name that doesn't exist at all is unambiguous and reported critical — verify the correct name before you or a teammate installs it. A name that does exist but was only published in the last 30 days is reported low/informational, not critical: this is routine for real new packages too, and on its own doesn't prove anything malicious — it's a "worth a second look" signal, not a verdict.
Features
- Checks npm and PyPI — package.json, requirements.txt, and pyproject.toml (both PEP 621
[project] dependencies and Poetry's [tool.poetry.dependencies])
- Direct dependencies only — fast, and matches where the actual slopsquatting risk moment is
- Scoped npm packages (
@scope/name) handled correctly — found and fixed a real URL-encoding bug during development (encodeURIComponent on the whole name double-encodes the registry's own @scope%2Fname convention, producing a URL the registry doesn't recognize as a single package)
- No API key required — queries the public, unauthenticated npm registry and PyPI JSON API directly
- SARIF 2.1.0 output — CI/CD integration (GitHub Code Scanning, etc.)
- 18 unit tests covering the manifest parsers (including a check that transitive lockfile entries are correctly not included) and the severity classification logic, all against real, deterministic fixtures — no mocked network calls, since parsing and classification are pure functions that don't need one. End-to-end verified against the real npm and PyPI registries with a fixture mixing real packages (
react, vitest, requests) and deliberately hallucinated ones — caught both hallucinated names, zero false positives on the real ones. Also run against this monorepo's own package.json files as a dogfooding check: zero false positives on real dependencies, and one genuinely correct low-severity finding (@guardbee/mcp-telemetry, itself only ~2 weeks old on npm at the time) — exactly the intended "worth a look, not an alarm" behavior.
Quick Start
Claude Desktop / MCP Client
CLI (CI/CD)
MCP Tools
Detected Patterns
Configuration (guardbee.yml)
Limitations (by design)
- Direct dependencies only. A hallucinated or squatted name inside a transitive dependency (something a package you depend on itself depends on) isn't checked — that package's own maintainers already resolved it against a real registry when they published.
- The "recently published" signal is genuinely ambiguous. Real new packages get published constantly; this is reported at low severity specifically because it is not proof of anything, only a "look closer" prompt.
- Network-dependent. Every check is a live HTTP request to the npm/PyPI registry — a project with many dependencies means many requests (bounded to 8 concurrent at a time), and a registry outage or rate limit will show up as a tool error, not a false "doesn't exist" finding (registry errors are reported separately from a genuine 404).
- No version-pin awareness. This checks whether the name exists, not whether the specific version range you've pinned is itself suspicious — that's a different, already-covered problem (see
@guardbee/mcp-dependency-auditor for known-CVE checking).
Development
License
MIT — GuardBee
Feedback
Questions, ideas, or a false positive? Start a thread in GitHub Discussions or open an issue. If GuardBee is useful to you, a ⭐ on the repository helps others find it.