pkgxray

pkgxray — pre-install security for npm packages, MCP servers, and AI agents

Inspect an npm package or MCP server before you install it or connect to it. You get a SAFE, REVIEW, or BLOCK verdict, decided by fixed rules and backed by cited evidence. The analysis is static and runs on your machine without installing npm dependencies. A pinned MIT-licensed Acorn parser is bundled. Normal scans never execute package code.

Website · Documentation · Calibration · Report a bug

Real runs: guard clears express@4.21.0 , then blocks a sample modeled on the 2024 @solana/web3.js compromise.

Highlights

• No npm dependency installation — local Node analysis with a bundled, pinned JavaScript parser.
• Install approved bytes — pkgxray install scans the npm lockfile, installs held archives offline with scripts disabled, and verifies the resulting files.
• Normal scans never execute package code — the tarball is read as bytes in quarantine.
• Cited verdicts from fixed rules — every finding names the file and the evidence. No model decides the verdict, so text planted in a package cannot steer it.
• Built for the agent era — check MCP servers before you connect, gate the installs an agent runs, and re-audit live MCP traffic.
• Adversarial gaps tracked openly — 200 additional synthetic cases track detected attacks and benign contrasts; this is internal testing, not an independent audit.
• Regression gates — calibration and adversarial corpora run in CI; historical top-1000 results remain available for comparison.

[1. Quick start](#quick-start) · 2. What it scans & detects · 3. Verdicts · 4. Usage · 5. Integrations · 6. How it compares · 7. Documentation

Why

AI coding assistants install packages and connect to MCP servers quickly, and often no person reads the code first. Sonatype counted more than 454,600 new malicious open-source packages across monitored ecosystems in 2025, over 99% of them on npm (Sonatype). npm audit asks whether a package has a known CVE. pkgxray also asks what the code does, before anything installs.

Quick start

1. Scan a known-benign package (no install of pkgxray needed):

npx --yes pkgxray@1.0.5 guard npm:express@4.21.0

It stages the tarball in quarantine and runs the static and supply-chain checks. There is no npm install, no lifecycle script, and no package code executed.

Sample output

Decision: SAFE   Grade: A+ (99/100)

No high- or medium-risk indicators were found in the provided evidence.

Notes:

• INFO npm-vs-github-clean — npm tarball matches the linked GitHub repo at the published version. (15/16 files match GitHub @4.21.0)

2. Read the verdict:

SAFE is not a proof that a package is harmless; static analysis cannot see a payload downloaded only at runtime. See the threat model.

3. See a BLOCK on the supplied inert fixture:

npx --yes pkgxray@1.0.5 --file examples/onboarding-malicious.json --format markdown

The fixture is inert source text that models a split-string SSH-key read and exfiltration. It is never executed. It returns BLOCK (exit 2) with the cited file and evidence.

4. Add it to your workflow — rechecks & CI, MCP, Hookshot install gate.

Two execution models. Default guard and audit scans are static, so

package code is never executed. Three surfaces are different: listing an MCP

server's tools may spawn it, mcp-proxy runs it behind a gate, and the opt-in

canary *executes* the package in a sandbox to

confirm what it does. The canary can confirm that a package is malicious, but it

can never prove one is safe. Full boundary: SECURITY.md.

What it scans & detects

Scans — pkgxray guard npm:name@version or pypi:name@version, github:owner/repo, a local directory, whole lockfiles across two ecosystems (npm: package-lock.json, yarn.lock, pnpm-lock.yaml, package.json; PyPI: requirements.txt, poetry.lock, Pipfile.lock, pyproject.toml), MCP servers, and AI-agent extensions.

Detects — credential theft (incl. split-fragment paths), cloud instance-metadata and secret-store harvesting, prompt injection, Unicode smuggling, base64 payloads and stage-2 loaders, exfiltration, persistence (shell profile, OS scheduler, and injected CI/CD workflows), self-deleting droppers, registry worm replication (install-time npm publish), npm install-hook and PyPI setup.py install-time execution, obfuscated computed-arg execution, hallucinated / slopsquat names (a lockfile pin the registry never published), known CVEs (via OSV, before download), npm↔GitHub artifact divergence, trojaned updates (recheck), and MCP capability-surface abuse.

The full coverage matrix is in the threat model, along with the known blind spot: a package that downloads its payload later. A side-by-side comparison table is on the website.

Verdicts

Exit codes are stable and CI-friendly: `0` safe/allow · `2` block · `3` review.

Usage

pkgxray guard npm:some-package@1.2.3 [--format json]   # vet a package before install

pkgxray guard pypi:some-package@1.2.3 # same, for a PyPI package (sdist staged + scanned) pkgxray mcp --package npm:some-mcp-server@1.4.2 npx some-mcp-server # vet an MCP server; --recheck catches the rug-pull pkgxray audit package-lock.json [--deep-all] # also: yarn.lock, pnpm-lock.yaml, package.json pkgxray audit requirements.txt [--deep] # PyPI: also poetry.lock, Pipfile.lock, pyproject.toml pkgxray recheck package-lock.json # scheduled: non-zero only on a regression

audit checks resolved dependencies against OSV. --deep adds source scans for blocked dependencies; --deep-all requests source scans for all resolved dependencies. Failed deep scans and incomplete source collection return REVIEW unless already BLOCK. guard --deps includes direct-dependency findings in its decision; only exact pins are checked, with ranges and other unresolved sources left at REVIEW. Guard output lists completed, partial, disabled and failed checks so a passing verdict does not imply checks that never ran. PyPI scans cover source-distribution manifests, vulnerability metadata, and text-level injection checks; they do not provide full Python module behavioral analysis or wheel inspection. Python source adds an unsupported-behavior REVIEW finding. Behavioral coverage gaps cannot be muted or promoted with allow-review; a pinned artifact approval remains an explicit operator override. Archives containing links or special files are rejected.

One optional .pkgxray.json tunes policy on the Node-based surfaces. The browser extension scans supplied evidence using engine defaults and cannot read project configuration. No config means the strictest settings. Config can never allow a CVE away, every loosening is printed, and a scan that errors fails closed to review. Schema and rules: configuration.md · .pkgxray.example.json.

Integrations

One engine behind every entry point. "Works with" means a documented setup guide, not a vendor-endorsed integration.

How it compares

npm audit and OSV-Scanner check for published CVEs, and pkgxray does not replace them. Run it alongside them. The tools in the same lane are Socket.dev, OpenSSF Package Analysis, and Cisco MCP Scanner, which also analyze what package code does. The full capability comparison is in docs/comparison.md and on the website.

Evidence

Historical top-1000 runs and their scope and methodology remain available at pkgxray.ca/stats. Those results predate the parser-based flow engine and have not been rerun for this change. The current engine passes all 270 internal synthetic calibration/challenge fixtures without a malicious SAFE or benign BLOCK; many malicious cases produce REVIEW. This is regression evidence, not a real-world detection-rate estimate.

Documentation

| compatibility.md · [json-schema.md]