Skills Janitor
Tinder for your Claude Code skills and MCP servers
338 followers
Tinder for your Claude Code skills and MCP servers
338 followers
v1.8, now in maintenance mode. claude code grew its own /skill-doctor and /doctor for spotting unused skills, so use those to trim context. janitor still does what isn't built in: a prompt-injection scan of every installed skill (and of one before you install it), duplicate detection, openai codex support, and actually deleting broken symlinks and dead skills. plus a swipe tui for skills and mcp servers. bash + python, zero deps. free, open source.
Skills Janitor
11% active rate sounds about right. I have a similar pile of skills I keep adding "just in case" and never trigger. The dedup feature is the real value here - I definitely have overlapping skills that quietly fight each other for context space. Going to run this today
Skills Janitor
@dklymentiev exactly same case for me. "Just in case" skills made my skills list too long and messy, like an inbox with all newsletter I signed up for but never opened.
Skills Janitor
@dklymentiev you were right about context space. /janitor-tokens now shows exactly how much each skill costs and which unused ones are eating budget. Also added codex support so it scans both platforms in one pass.
Okan
I've been dumping so many custom tools into my Claude Code config lately that half my context window is probably just unused skill descriptions. I am really curious if this parses the local Claude logs to calculate usage stats or if I need to run it alongside a proxy. Hooking this up to an automated weekly cleanup script would be a massive time saver.
Skills Janitor
@y_taka it reads your local conversation history (~/.claude-account-personal/history.jsonl), no proxy needed. it matches slash commands directly and also does keyword similarity matching for when you reference skills without the slash command.
you get explicit vs estimated usage counts with weekly granularity. the context window bloat thing is real though, that's actuallyone of the features i want to add next - showing how many tokens each skill's system prompt eats so you can see the actual cost of keeping unused ones around.
for the weekly cleanup - right now you'd just cron /janitor-report and pipe the output, but an auto-clean mode with a usage threshold is on the roadmap too.
Skills Janitor
@y_taka v1.1 adds /janitor-tokens which shows the actual token cost per skill. still reads local history, no proxy needed. for weekly cleanup you can cron /janitor-report and pipe the output, auto-clean mode with a usage threshold is next on the list
Skills Janitor
@klashkil yes, janitor-duplicates is showing overlap % for similar skills. example below:
@krzysztof_hendzel oh this is great, i would prolly merge a bunch of skills together xD
Can I also use this tool as a pre-check before installing a new skill to see if the same or a similar skill is already present, maybe with a different name?
Skills Janitor
@chintant hmm good call, I will probably add it in the next version. It's better to check skills before even installing, keeping your setup clean is my top goal.
Skills Janitor
@chintant shipped it. /janitor-precheck does exactly this now — give it a github url and it checks overlap against everything you have installed before you add it. thanks for the push on this one :)
This solves a problem I didn't realize I had until I saw it. I've been building with Claude Code daily and my skills folder is a mess duplicates, things I set up once and forgot about, broken ones I never cleaned up. The audit plus dedup combo is exactly right. Love that it's open source too. What was the hardest part detecting overlapping skills that use different wording?
Skills Janitor
@linoy_bar_gal honestly yes, that was the part that took the longest. keyword matching missed too much (same skill, different verbs), pure embeddings missed too little (flagged unrelated skills with similar tone). ended up with a hybrid: token-level jaccard on descriptions + name similarity, weighted. not perfect but it catches the obvious 60% and stays explainable.
solid util. though if you're surprised by which skills you actually use, the taxonomy was probably wrong to begin with.
Skills Janitor
@mykola_kondratiuk could be, I'm thinking about making it more simple or make it more like a set of actions under fewer commands, will see. thanks for the comment :)
Skills Janitor
@mykola_kondratiuk took your point about the taxonomy. consolidated from 9 to 7 commands -check/duplicates/cleanup/compare all merged into report/fix/search. Fewer things to remember, same coverage :)