Stack Cleaner

See, organize, and clean up your Claude Code setup: every skill, plugin, MCP server, and agent, split by global vs. project.

Free · open source · runs locally · sends nothing.

Live app →  ·  Guided setup →  ·  Try the demo →

Over time, Claude Code accumulates a lot: user-level skills in ~/.claude, project skills in every repo's .claude, plugins from a handful of marketplaces, MCP servers wired into different projects, and a pile of sub-agents. It gets hard to answer simple questions: what do I actually have? what do I never use? what's duplicated? what can I safely remove?

This tool answers them. You run one command, it reads your local Claude Code install, and the web app shows everything grouped by where it lives (global vs. each project), with real usage counts and automatic duplicate detection, so you can spot the dead weight, see what's redundant, and build a one-click cleanup plan.

How it works

  ┌─────────────┐      ┌──────────────────────┐      ┌─────────────────────┐
  │  1. Scan    │  →   │  2. See              │  →   │  3. Tidy            │
  │  one line   │      │  global vs project   │      │  cleanup plan, or   │
  │  in terminal│      │  + real usage counts │      │  hand it to Claude  │
  └─────────────┘      └──────────────────────┘      └─────────────────────┘

1. Scan: run the one-liner below. It writes stack-cleaner.json next to you. No dependencies, nothing sent anywhere.
2. See: drop that file into the web app. It's parsed in your browser; nothing is uploaded.
3. Tidy: tick the items you want gone and export a shell script, a paste-to-Claude prompt, or JSON.

The scan

The one command, same on macOS, Windows, and Linux, no curl, nothing piped:

npx stack-cleaner@latest

It runs the published npm package (which bundles the same scanner) and writes stack-cleaner.json next to you. Node ships with Claude Code, so you already have it.

macOS / Linux (bash/zsh):

curl -fsSL https://stackcleaner.com/scan.mjs | node

Windows (PowerShell): curl there is an alias for Invoke-WebRequest, so call curl.exe and download-then-run instead of piping:

curl.exe -fsSL https://stackcleaner.com/scan.mjs -o stack-cleaner-scan.mjs; node stack-cleaner-scan.mjs

Prefer to read it first? It's one short, dependency-free file (public/scan.mjs). Download then run (use curl.exe on Windows):

curl -fsSL https://stackcleaner.com/scan.mjs -o stack-cleaner-scan.mjs
node stack-cleaner-scan.mjs

Not comfortable in a terminal? The Setup page walks you through it copy-paste by copy-paste. You never need a GitHub account or any coding.

What it captures

For every item it records the type, the scope (global or which project), a description, and a real usage count: how many times you've actually invoked it, and when you last did.

| Type | Global source | Project source | |------|---------------|----------------| | Skills | ~/.claude/skills/*/SKILL.md | <repo>/.claude/skills/*/SKILL.md | | Plugins | ~/.claude/plugins/installed_plugins.json | – | | MCP servers | ~/.claude.json → mcpServers | per-project mcpServers + .mcp.json | | Agents | ~/.claude/agents/*.md | <repo>/.claude/agents/*.md |

It enumerates every project Claude Code knows about (from ~/.claude.json), not just the folder you run it in, so you get your whole picture in one pass.

Where usage counts come from

The scan reads your local Claude Code transcripts (~/.claude/projects/*.jsonl) to count real invocations for skills, agents, and MCP servers, so you can finally see what's installed but never used, even for the types that carry no usage count in plain config. It extracts only the tool / skill / agent / MCP-server names, the counts, and the timestamps, never your prompts, message text, tool arguments, file paths, or command contents. It all stays local until you choose to upload the file. Pass --no-transcripts to skip the transcript read entirely. (Skills and plugins also keep their counts from Claude Code's own skillUsage / pluginUsage tables.)

Finding duplicates

Claude Code makes it easy to end up with the same thing twice — a standalone skill you installed by hand that a plugin now bundles, the same MCP server wired up from two sources, or same-named copies across global and a project. Stack Cleaner flags these automatically:

• Superseded by a plugin — a standalone skill, agent, or MCP server that an installed plugin already provides. The plugin's copy stays updated; the loose one won't, so it's the safe one to remove.
• Same-name duplicates and duplicate MCP servers — the same item installed two different ways.

Every flagged item explains, in plain language, what it duplicates, where your copy lives, and which one to keep. A Duplicates filter and a "Select N redundant copies" button turn the whole pile into a one-click cleanup.

Privacy & safety

This is the part that matters, because the output describes your tooling.

• Runs entirely on your machine. The scan never makes a network request. The web app parses your file in the browser and stores it only in that browser's localStorage. It is never uploaded.
• Transcripts: names and counts only. To compute real usage it streams your local transcripts but extracts only tool/skill/agent/MCP-server names, counts, and timestamps, never prompt text, arguments, file paths, or command contents. Opt out with --no-transcripts.
• Secrets are stripped before the file is written. MCP env values, auth headers, URL credentials, and token-looking arguments are replaced with <redacted>. Your home directory is rewritten to ~ so your username doesn't leak.
• It only reads. The scan never modifies, installs, or removes anything. Removal commands are generated as text for you to review and run.
• No dependencies, no build step to scan. scan.mjs is plain Node with node: built-ins only. Read the whole thing in a minute.

The cleanup plan

Select items and the tool builds three things:

• Hand to Claude: a prompt you paste back into Claude Code so it does the removals for you.
• Shell script: the exact claude plugins uninstall … / rm -rf ~/.claude/skills/… / claude mcp remove … commands, grouped by type, for you to review and run.
• JSON: a machine-readable selection for your own tooling.

Nothing is ever removed by this tool. You stay in control.

Run it yourself / self-host

It's a standard Next.js (App Router) app with no backend and no database.

git clone https://github.com/BespokeWoodcraftStudio/stack-cleaner
cd stack-cleaner
npm install
npm run dev        # http://localhost:3000

# scan your own machine straight from the source:
node public/scan.mjs

Or install the published CLI globally and run it from anywhere:

npm i -g stack-cleaner
stack-cleaner          # writes stack-cleaner.json in the current folder

Deploy your own copy to Vercel (or any static-capable Next host) with one click of "Import Project". There's nothing to configure.

The inventory schema

The scan emits, and the app reads, a single JSON shape, abbreviated below; the full field list is in lib/types.ts:

{
  "schemaVersion": 1,
  "generatedAt": "2026-…",
  "generator": "[email protected]",
  "machine": { "platform": "darwin", "node": "v22.0.0" },
  "projects": ["my-app", "my-blog"],
  "usageSummary": {             // present when the transcript scan ran
    "totalInvocations": 1280,
    "itemsWithUsage": 24,
    "itemsUnused": 11,
    "transcriptsScanned": 4142,
    "generatedFrom": "transcripts"
  },
  "items": [
    {
      "id": "skill:global:graphify",
      "type": "skill",            // skill | plugin | mcp | agent
      "scope": "global",          // global | project
      "project": null,            // basename for project-scoped items
      "name": "graphify",
      "description": "…",
      "path": "~/.claude/skills/graphify",
      "usageCount": 10,           // mirrors invocationCount when matched
      "lastUsedAt": 1718000000000,
      "usageClass": "good",       // good | warn | bad | info | unknown
      "usageLabel": "✅ 10 uses", // display string
      "invocationCount": 10,      // transcript invocations (0 = tracked, never used)
      "lastUsed": 1718000000000,  // epoch ms of most recent invocation, or null
      "usageSource": "transcripts", // "transcripts" | "claude-json" | "none"
      "removeCmd": "rm -rf ~/.claude/skills/graphify"
    }
  ]
}

Because it's just JSON, you can hand-write or generate an inventory from any source and drop it in.

Project layout

public/scan.mjs              the local scanner (zero-dep Node)
lib/types.ts                 the shared inventory schema
lib/inventory.ts             parse / filter / group / manifest logic
lib/demo.ts                  the built-in demo dataset
app/page.tsx                 landing
app/setup/page.tsx           guided, non-technical walkthrough
app/inventory/page.tsx       the tool
components/                  ui primitives + page sections

Documentation

• FAQ (or the in-app /faq): is anything uploaded, are my keys safe, can it delete things, Windows support…
• Usage guide: a walkthrough of the inventory tool: filtering, selection, and the three cleanup exports.
• Setup wizard: the non-technical, copy-paste path.
• Security policy: how to privately report a redaction miss (and what we already do).
• Support / troubleshooting: common errors and where to get help.
• Changelog · Contributing · Code of Conduct

Contributing

Issues and PRs welcome. See CONTRIBUTING.md for the dev setup and the verify gate, and SECURITY.md for reporting a secret leak privately. Good first additions: more cleanup-command coverage, richer usage-trend views, and broader plugin-marketplace detection.

License

MIT, see LICENSE. A Bespoke Woodcraft Studio tool. Not affiliated with Anthropic; "Claude" and "Claude Code" are trademarks of Anthropic.