safedeps

Stop your AI coding agent from installing vulnerable or unapproved dependencies — and roll back the ones that slip through.

safedeps gates every dependency install your Claude Code or Codex CLI agent runs. It pre-approves packages against OSV / CISA KEV / GitHub Advisory, re-verifies the closure that actually lands in your lockfile, and auto-rolls-back anything that diverges. Local-only, with zero runtime dependencies. (한국어 README → README.ko.md)

• Pre-approve — every pkg@version, plus its full transitive closure for npm, is cleared against OSV (canonical), CISA KEV, and GitHub Advisory before it installs.
• Enforce the real effect — after the install, the actual package-lock.json closure is re-checked, so a wrapped or obfuscated command can't sneak a package past the gate.
• Roll back — anything unapproved or newly-vulnerable is reverted to the last confirmed safe snapshot, or to the state before the command when a project has none yet, and the message says which. On Claude Code safedeps adds --ignore-scripts to each npm install, so that a rejected package's lifecycle scripts do not run during the install. That is the design, not a promise: whether npm keeps the flag is decided by shell state the command does not show, so safedeps reports only that it added the flag.

A real catch. The pre-commit audit flagged a vulnerable transitive hono advisory that Dependabot missed — by re-querying the advisory DB at commit time. A CVE disclosed after you installed a package ("looked safe then, flagged now") surfaces at your next commit, not weeks later.

Quickstart

# 1. Install the CLI — the npm package is scoped, note the @aldegad/ prefix
npm install -g @aldegad/safedeps

# 2. Wire the hooks into Claude Code / Codex (idempotent)
cd "$(npm root -g)/@aldegad/safedeps" && node scripts/install/install-safedeps-hooks.mjs

# 3. Done — every dependency install your agent runs is now gated.

safedeps is the CLI command; the npm package is @aldegad/safedeps — the unscoped safedeps on npm is an unrelated package. Prefer the full skill source tree? See Installation.

The hooks are one program, safedeps-core, written in Rust. The npm package carries it built, for macOS (Apple silicon and Intel) and Linux x64, and the installer runs it once before it registers anything. A checkout builds it with cargo instead. Where the binary comes from, and what you see when it is missing or out of date, is under Where the hook binary comes from. What v2.19.0 changed for you is under What Changed in v2.19.0.

Distribution Model

Safedeps has two distribution surfaces:

1. Agent skill + hooks (canonical) -- the repo itself is the skill folder. SKILL.md, the hook entry and its Rust core, provider/ledger libraries, and install helpers stay together in one directory. A checkout holds the core's source and builds its binary; the npm package carries the built binaries.
2. npm package (CLI convenience) -- @aldegad/safedeps installs the safedeps command. npm does not make Claude Code or Codex automatically discover the skill; after npm installation, users still need to run the hook/skill installer or manually register the skill folder.