# Skill Cabinet Skill Cabinet is an open source (MIT) local web app that shows you every [[AI Agent Skills|agent skill]] installed on your machine, in one place. One command starts it: `npx skill-cabinet` (Node 20+). It runs a small server bound to `127.0.0.1` (port 3781 by default) and opens the catalog in your browser. Nothing leaves your machine. Why does this matter? Because skills pile up fast. You install a few from [[Skills.sh]], a plugin drags in ten more, [[Claude Code]] has its own folder, [[Codex CLI]] another one, [[Cursor]] yet another... After a few months, you have no idea what's actually loaded, which copies are duplicates, and which ones point to nothing. I've been there: with hundreds of skills in my own setup, "what do I have installed?" is NOT a trivial question anymore. ## What it does Skill Cabinet scans the "drawers" where agents look for skills: `~/.agents`, `~/.claude`, `~/.codex`, `~/.cursor` (plugins included), Hermes profiles, and any other `~/.*/skills` folder. For each skill, you can: - Read the body (rendered or source), the YAML frontmatter and any bundled files - See whether it's a folder, a single file or a symlink - Follow its GitHub origin when the skill declares it or the install path encodes it (origins guessed from a parent plugin or git remote are flagged as inferred) - Spot identical copies across drawers, virtual references and broken links - Get a static risk flag on the skill body - Filter by drawer, form (physical, reference, broken), risk and invocation mode (user-only, model, hook, off), and search across name, description, path, origin and frontmatter There's also a "house census" view: how much disk the skills take, what's real vs what merely points somewhere, and how many bytes the duplicates waste. ## Quarantine instead of delete This is the feature I like most. Quarantine moves a skill out of every drawer agents read (into `~/.skill-cabinet/quarantine`) without destroying it. Restore puts it back. That's exactly what you want when you suspect a skill is polluting your context or triggering at the wrong time: pull it out, see if things improve, bring it back if not. Delete is permanent (no undo). Deleting a symlink only removes the link, the target stays. Note that built-in [[Cursor]] skills and plugin cache copies can reappear the next time the tool updates. Keyboard driven: `j`/`k` to move, `/` to search, `x` to mark, `q` quarantine, `r` restore, `d` delete. ## Why I care Skills are becoming the unit of work for [[AI Agents]], and like any collection, they need curation. More skills isn't better: every installed skill costs a bit of context and adds a chance of the wrong one firing. Skill Cabinet gives you the audit trail that the agent tools themselves don't (yet) provide. In my own vault, I solved the discovery side differently, by serving skills on demand through the [[Agentic Resource Discovery Server plugin for Obsidian]] instead of preloading them. The two approaches are complementary: one decides what agents *see*, the other shows you what's *on disk*. ## References - https://github.com/subsy/skill-cabinet ## Related - [[AI Agent Skills]] - [[Claude Code Skills]] - [[Skills.sh]] - [[My Published AI Skills]]