plugin-auditor

plugin-auditor is a Claude Code plugin that performs a security audit of other Claude Code extensions, such as skills, agents, hooks, MCP servers, and slash commands, before a user installs them. The README explains that this kind of ecosystem, where anyone can publish a plugin that a single command will clone and install, is a classic supply chain risk: a malicious skill could try to leak credentials, a hook could persist across sessions, an MCP server could fetch and run arbitrary code, or an install script could run before anyone reviews it.

The tool works entirely by reading and analyzing code; it never runs or executes anything from the repository being audited. It adds one slash command, /plugin-auditor:audit, which is not triggered automatically, a user has to run it on purpose. When run, it hands the work to five specialized sub-agents that each look at a different risk area, covering static code patterns, Claude specific artifacts, supply chain concerns, configuration files, and suspicious network or filesystem activity. Their findings are combined into a single markdown report with an overall verdict of SAFE, CAUTION, or UNSAFE, and every finding is tied to a specific file and line number so a user can check it themselves.

The audit can be pointed at the current folder, an explicit local path, or a public GitHub or GitLab URL, which the tool will shallow clone into a temporary folder before scanning. There is also a delta mode that only audits what changed since a repository's last audit, using a saved state file to compare against the previous commit. Reports are written to a folder under the user's home directory and are timestamped, so re-running an audit never overwrites an earlier report. After a report is produced, the tool can also walk through any concerning findings interactively, one section at a time.

It is written mainly in shell script, is licensed under MIT, and is intended for people who install community-made Claude Code plugins and want a way to check them for hidden risks before trusting them.