Audit every skill, subagent and MCP server your coding agent loads, across Claude Code and OpenAI Codex. 6 commands, zero dependencies.
Skills Janitor finds what you have never actually used (counted from real session transcripts, including skills the agent auto-triggered), what duplicates something you already had, what is broken, and which connected MCP servers load their tool schemas on every request for nothing.
It is honest about the cost. Only a skill's description sits permanently in your system prompt; the body loads when it triggers. Both numbers are reported separately, instead of quoting the whole file as context rent.
Then it deletes. /janitor-swipe turns the collection into a Tinder deck sorted heaviest-and-least-used first: swipe left to remove, right to keep, with timestamped backups so every removal is reversible. /janitor-security scans for prompt injection and hidden instructions, and /janitor-discover checks overlap and safety before you install anything new.
The alternatives are read-only. They tell you what is bloating your context; Skills Janitor cleans it. Bash and Python 3, no npm or pip install, MIT licensed.
Ranked #4 Product of the Day on Product Hunt with 250+ upvotes.
