Pkgxray - inspect what gets installed, not what executes

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

Use local, zero-dependency package static analysis to inspect npm packages and Model Context Protocol (MCP) servers before installation or connection. pkgxray reports cited SAFE, REVIEW, or BLOCK evidence without executing package code during normal scans.

• Static analysis
• Supply-chain intelligence
• Prompt-injection detection
• MCP security
• SAFE / REVIEW / BLOCK

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

60-second walkthrough

npx --yes pkgxray@1.0.4 guard npm:express@4.21.0

This downloads pkgxray through npm's temporary npx cache, stages the target tarball in quarantine, and performs the static and supply-chain checks. It does not globally install pkgxray, run npm install, execute lifecycle scripts, or execute package code.

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) … Real output, abridged.

A BLOCK verdict instead lists every finding with the file and evidence that produced it.

Point it at a package, get a verdict with cited evidence - before a single line of that package runs. guard stages the package in a sandboxed quarantine, audits the staged copy, and only promotes it when policy allows. It never runs npm install, lifecycle scripts, build steps, or package code.

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

From a repository checkout:

npx --yes pkgxray@1.0.4 --file examples/onboarding-malicious.json --format markdown

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

• Scan pull requests and schedule dependency rechecks.
• Expose pkgxray's tools to an MCP-capable coding agent.
• Evaluate the experimental Hookshot install gate.

Why pkgxray exists

AI coding assistants install packages and connect to MCP servers at machine speed, often without a human ever reading the code. Sonatype reported 454,648 newly identified malicious open-source packages across monitored ecosystems in 2025. Its Q4 report counted 394,877 in that quarter and said 99.8% of Q4 malware originated from npm (annual figure; Q4 scope).

Traditional antivirus inspects what executes; pkgxray inspects what gets installed. npm audit and OSV-Scanner answer an essential question - does this package have a known CVE? - and pkgxray asks it too (via OSV, before anything downloads). But a freshly trojaned package has no CVE yet, so pkgxray also analyzes trust: what the code actually does, whether the published npm artifact matches the tagged GitHub source, whether the provenance attestation is consistent with the claimed repository, and whether the docs carry a prompt-injection payload aimed at the agent reading them.

It is intentionally conservative: verdicts come from deterministic heuristics (no LLM in the verdict path, so injected text can't steer them), only citable evidence is reported, and the zero-heuristic-false-block calibration on the top-1000 most-downloaded packages is regression-gated in CI. That claim is scoped to the most-installed set - it is not a claim of zero false blocks on every package; the newer MCP/agent-tooling ecosystem is over-blocked and being reconciled per-case (details).

Threat coverage

✅ detected · ◑ partial / indirect

Known blind spot: pkgxray reasons about bytes in the tarball. A package that downloads its real payload after install can ship a clean tree - pkgxray flags the capability when its shape is unambiguous, but pair it with runtime sandboxing when that risk matters.

Full analysis: docs/threat-model.md.

Surfaces

• Continuous monitoring - pkgxray recheck diffs installed deps against a stored verdict baseline and pre-vets newer versions
• MCP vetting - pkgxray mcp audits a server's tool manifest before you connect; --pin / --recheck catch the rug-pull; pkgxray-mcp gives any agent the audit tools directly
• Runtime gate - pkgxray mcp-proxy wraps a live MCP server on the wire: denied tools stripped, ~0.05 µs per-call verdict, injection scan of tool results
• Install gate - a hookshot hook runs guard on every package an agent tries to install, across Claude Code, Cursor, Windsurf, Factory Droid, and Codex (examples/hookshot/)
• Policy engine - one .pkgxray.json read by every surface; tighten freely, every loosening is printed; CVEs can never be allowed away; fail closed
• Opt-in behavioral canary - pkgxray canary runs lifecycle scripts in an OS sandbox with decoy credentials; it can confirm malice, never clear a package

Verdict reference

Exit codes are stable and CI-friendly: 0 safe/allow · 2 block · 3 review. The full signal-to-severity mapping is in the severity policy.

Commands

Vet an npm package before installing

pkgxray guard npm:some-package@1.2.3 [--format json]
pkgxray guard ./ext --promote-to ./approved/ext   # local dir, promote if policy allows

Vet an MCP server before connecting - full guide: docs/mcp.md

pkgxray mcp --package npm:some-mcp-server@1.4.2 npx some-mcp-server
pkgxray mcp --recheck npx some-mcp-server   # catch the rug-pull

Enforce in CI/CD

pkgxray audit package-lock.json [--deep]   # also: yarn.lock, pnpm-lock.yaml, package.json
npx pkgxray recheck package-lock.json       # scheduled: exits non-zero only on a regression

A ready-made GitHub Actions integration and the self-hostable cache server (PKGXRAY_CACHE_URL) are documented in the reference.

Guard AI coding agents

pkgxray is published on the MCP Registry as io.github.adamsjack711-ux/pkgxray. Add it to any MCP client - locally installed (pkgxray-mcp) or zero-install via npx:

{
  "mcpServers": {
    "pkgxray": {
      "command": "npx",
      "args": ["--yes", "--package", "pkgxray@1.0.4", "pkgxray-mcp"],
      "env": {
        "PKGXRAY_MCP_ALLOWED_ROOTS": "/absolute/path/to/project"
      }
    }
  }
}

The MCP guide explains the operator-owned filesystem boundary. Product-specific setup is in the coding-agent integration guide.

Gate installs with the Hookshot integration and wrap MCP servers with pkgxray mcp-proxy. One optional .pkgxray.json, read by every surface. Zero config means maximum strictness. Precedence, mute / mcp blocks, and enforced invariants: docs/configuration.md · .pkgxray.example.json.

Walkthrough video

The 60-second walkthrough - the SAFE run, the blocked trojan with its exit code, then a lockfile audit: pkgxray-demo.mp4

All captures are real runs - reproduction steps in docs/screenshots/, which also shows the MCP proxy, hookshot install gate, and browser extension in action.

Comparison with other tools

npm audit and OSV-Scanner match dependencies against known CVEs - a different question, answered well. pkgxray is designed to run alongside them, not replace them (it queries OSV itself, before anything downloads). The comparison that matters is against tools in the same lane - behavioral supply-chain vetting:

Comparison made 2026-07-21 against each tool's public documentation; unknown means not publicly documented - not verified either way.

¹ Socket's analysis runs in its cloud; Socket Firewall needs no account but consults Socket's hosted intelligence on every install.
² Open source and self-hostable, but built as a registry-scale analysis pipeline (Docker/gVisor), not an install-time developer gate.
³ The YARA analyzer runs locally; the LLM-as-judge and Cisco AI Defense analyzers require API keys.
⁴ Both detonate packages in an OS sandbox. pkgxray's opt-in canary runs two phases - install-time lifecycle scripts and the import of the package entry point - with decoy credentials, so the malicious-on-first-require (flatmap-stream) shape that is pkgxray's stated blind spot is triggered and observed. Egress is now kernel-confined on both platforms: sandbox-exec on macOS and, on Linux with bubblewrap + iproute2, a private network namespace (bwrap+netns) where a raw-socket dial that bypasses the proxy is refused by the kernel (ENETUNREACH) while proxied egress is still captured. That tier engages only after a runtime self-test proves it in the environment (verify with node scripts/verify-netns-confinement.js); absent the tooling it falls back to observe-only and says so. Still ◑ - not for a confinement gap, but by design: canary is opt-in and confirm-only (it proves malice, never clears a package) and detonates without the package's dependencies installed, whereas OpenSSF Package Analysis runs registry-scale and default-on. Run them as complements - pkgxray before install, full dynamic analysis where that risk matters.
⁵ Socket's LLM-based code inspection is a headline feature (“AI-detected potential malware”, human-confirmed); Cisco's YARA-only mode is deterministic, its LLM analyzer is not.
⁶ Socket Firewall blocks risky packages at install time; it does not stage a quarantined copy for human review.
⁷ Socket's MCP offering exposes its package-scoring API to agents; it does not vet arbitrary MCP servers at connect time.
⁸ Cisco MCP Scanner is analysis-only per its docs - it does not proxy or gate live MCP traffic.

Architecture

Acquisition (OSV pre-check → fetch) → sandboxed quarantine → static analysis → policy → verdict. The same engine backs every surface: CLI, MCP server, runtime proxy, install hook, browser extension, and CI cache server.

Principles: never execute untrusted code · citable evidence only · minimize false positives · fail closed · zero runtime dependencies.

Details: docs/architecture.md · docs/design.md

Performance

• Local static analysis: ~25 ms - a full guard of express is ~1.3-1.5 s cold-cache, almost all network round-trips (Apple M1, Node 26)
• Known-vulnerable packages block at the OSV pre-check, before download
• Calibration (precision, recall, the 0-heuristic-false-block gate on the top-1000 most-downloaded - scope) is measured by a committed benchmark corpus that fails CI when it regresses

Full numbers: docs/reference.md#performance · methodology: docs/benchmark.md

Documentation

Start at the documentation index.

Future

Longer-term plans: project status, adoption playbook, and GitHub issues.

npm test   # zero-dep node --test suite
npm run benchmark   # calibration corpus: precision/recall + 0-false-block gate