guard

Guard

Supply chain security for pnpm repositories and GitHub Actions.

Guard helps teams answer one question before a dependency or CI change lands:

Can we trust this change?

It combines:

• repository posture checks,
• dependency and lockfile review,
• workflow hardening,
• policy validation and exceptions,
• machine-readable output for CI and automation.

Links:

• Changelog
• Releases

Install

Target	Install
Claude Code	claude plugins marketplace add MauroProto/guard && claude plugins install guard@guard
Guard CLI (Go)	go install github.com/MauroProto/guard/cmd/guard@latest
Guard CLI (shell)	curl -fsSL https://raw.githubusercontent.com/MauroProto/guard/main/install.sh | sh
From source	git clone https://github.com/MauroProto/guard.git && cd guard && make install

If the Claude Code plugin cannot find the guard binary in PATH, set:

export GUARD_BIN=/absolute/path/to/guard

What Guard ships

Guard is a hybrid product:

• Guard CLI stays first-class for CI, SARIF, JSON, headless scans, review-pr, baseline, and policy lint.
• Claude Code plugin adds focused, contextual Guard runs inside edit sessions without replacing the CLI.

The plugin is distributed from this repository through .claude-plugin/marketplace.json, and the Claude Code plugin itself lives at:

plugins/claude-code/guard-security

Quick start

guard init
guard scan
guard fix
guard baseline record

For a PR or branch review:

guard review-pr
guard review-pr --base origin/main --head HEAD --format markdown
guard explain review.diff.install_script.added

For focused scans:

guard scan --scope workflows --format json
guard scan --scope deps --files package.json,pnpm-lock.yaml --format json
guard scan --scope policy --files .guard/policy.yaml --format json
guard scan --changed-files --format json

Claude Code plugin

The plugin is designed to be useful, not noisy.

When enabled, it reacts to the right moments:

• SessionStart
  • verifies Guard availability,
  • detects repo root and key files,
  • initializes light session state.
• FileChanged
  • marks deps, workflows, workspace, or policy as pending review.
• PostToolUse on Write|Edit
  • runs focused Guard scans in JSON mode for affected surfaces.
• PreToolUse on Bash
  • watches dependency mutation commands such as pnpm add, pnpm up, pnpm install, npm install, and corepack use.
• Stop
  • summarizes pending or blocking Guard scopes.

V1 behavior is intentionally balanced:

• no full scans on every event,
• no auto-fix from hooks,
• no blanket Bash blocking,
• no dependence on paths outside the installed plugin bundle.

One-off local plugin testing

claude --plugin-dir /absolute/path/to/guard/plugins/claude-code/guard-security

Commands

Command	Alias	Description
guard scan	guard s	Scan the repository for security issues
guard fix	guard f	Apply safe local remediations
guard init	guard i	Create or patch a secure baseline
guard ci	guard c	Strict scan mode for CI pipelines
guard diff	guard d	Compare two package versions for risk signals
guard review-pr	guard review	Review dependency and workflow changes between git refs
guard approve	guard ab	Approve a package that needs build scripts
guard baseline record	-	Record the current finding set as baseline debt
guard explain	-	Explain a rule ID or a finding fingerprint
guard policy lint	-	Validate policy/config semantics and deprecated fields

What it checks

Repository posture

• missing pnpm-lock.yaml
• missing packageManager in package.json
• missing engines.node

pnpm hardening

• minimumReleaseAge missing or too low
• blockExoticSubdeps disabled
• strictDepBuilds disabled
• trustPolicy not set to no-downgrade
• unapproved build scripts in allowBuilds

GitHub Actions

• actions not pinned to full commit SHA
• missing or overly broad permissions
• missing CODEOWNERS
• risky pull_request_target patterns
• privileged publish workflows without attestations

Dependency review

• new install scripts
• suspicious remote fetch or command execution patterns
• new binary files
• obfuscation signals
• sensitive path access like .env, .ssh, .npmrc
• new OSV advisories
• trust regressions surfaced by review-pr

Output formats