ithiria894

Posted on Mar 23 • Edited on Apr 15 • Originally published at github.com

Claude Code Secretly Hoards 140+ Config Files Behind Your Back. Here's How to See What It Actually Loads.

#ai #opensource #mcp #showdev

Fair warning: I'm a CS dropout. I'd been using Claude Code for less than two weeks when I peeked inside ~/.claude/ and found 140 items I never knew existed — memories, skills, MCP server configs, commands, agents, rules, hooks, plans, sessions — scattered across directories. I built a dashboard to fix it in one day. Within weeks it had picked up 250+ GitHub stars and users from 15+ countries.

Here's the problem I found, and the tool I built to solve it.

You Don't Know What Claude Actually Loads

Every time you start a Claude Code session, it silently loads configs from multiple locations. But here's what nobody tells you: each category follows different rules.

MCP servers use local > project > user precedence — same-name servers in a narrower scope override broader ones
Agents — project-level agents shadow same-name user agents
Commands — available from user and project, but same-name conflicts are officially "not supported"
Skills — available from personal, project, and plugin sources
Settings — resolved by a precedence chain: managed > CLI > project local > project shared > user
CLAUDE.md — walks up parent directories from your working directory

There is no built-in command to see what's effective in your current session. You have to mentally reconstruct it from multiple directories, JSON files, and precedence rules.

The best part: I found three identical MCP server entries because I'd added the same server while cd'd into different directories. Claude happily duplicated it each time without telling me.

Why This Actually Hurts Your AI's Performance

This isn't just messy — it's a technical problem. Everything Claude loads goes into your context window.

A skill meant for a Python pipeline sitting in global? It gets loaded into every session — including your React frontend work. Duplicate MCP entries? Claude initializes the same server multiple times. Stale memories from months ago? Still there, still eating tokens, still potentially contradicting your current instructions.

Irrelevant configs directly degrade LLM accuracy. Your context window is finite — every byte wasted on misplaced config is a byte not spent on your actual work.

One Command, Full Visibility

npx @mcpware/claude-code-organizer

Opens a web dashboard at http://localhost:3847. Scans your real ~/.claude/ directory. Shows everything Claude Code has stored — across all projects.

Show Effective — See What Claude Actually Loads

This is the core feature. Click ✦ Show Effective and CCO computes what Claude actually loads for each project, using each category's own official rules.

Items are tagged with status badges:

GLOBAL — available from ~/.claude/, applies to all projects
ANCESTOR — loaded from a parent directory (CLAUDE.md ancestry)
SHADOWED — overridden by a project-level item with the same name
⚠ CONFLICT — same name exists in multiple scopes with unclear precedence

Hover any category pill for its specific rule. Categories without official rules (plans, sessions) are dimmed — CCO doesn't fake rules that don't exist.

What You Actually See

Category	Viewable	Movable	Deletable
Memories (feedback, user, project, reference)	Yes	Yes	Yes
Skills (with bundle detection)	Yes	Yes	Yes
MCP Servers	Yes	Yes	Yes
Commands (slash commands)	Yes	Yes	Yes
Agents (subagents)	Yes	Yes	Yes
Rules	Yes	—	Yes
Plans	Yes	—	Yes
Sessions	Yes	—	Yes
Config (CLAUDE.md, settings.json)	Yes	Locked	—
Hooks	Yes	Locked	—
Plugins	Yes	Locked	—

11 categories across all projects — not 15 rounds of cat and ls.

Not Just a Viewer

Move items where they belong

Memory trapped in a project that should be global? Click Move, pick the destination. MCP server that belongs to one repo? Move it there. Warnings are shown for precedence changes and name conflicts before any move happens.

Context Budget

Your context window is not 200K tokens. It's 200K minus everything Claude pre-loads.

CCO counts every item's tokens (ai-tokenizer, ~99.8% accuracy), shows always-loaded vs deferred breakdown, expands @imports to see what CLAUDE.md actually pulls in, and lets you toggle between 200K and 1M context windows.

MCP Security Scanner

Every MCP server you install exposes tool descriptions that go straight into Claude's prompt. A compromised server can embed hidden instructions you'd never see.

CCO connects to every MCP server, retrieves actual tool definitions, and runs them through:

60 detection patterns cherry-picked from 36 open source scanners
9 deobfuscation techniques (zero-width chars, unicode tricks, base64, leetspeak)
SHA256 hash baselines — if a server's tools change between scans, you see a CHANGED badge immediately

Other features

Flat or Tree view — Default flat view lists all projects equally. Toggle tree view (🌲) to inspect filesystem structure
Find duplicates — All items grouped by category. Three copies of the same memory? Delete the extras
Undo everything — Every move and delete has an undo button
Bulk operations — Select mode: tick multiple items, move or delete all at once
Detail panel — Click any item to see metadata, content preview, and a "Why it applies" explanation
Search & filter — Real-time search across names, descriptions, and paths
MCP tools — Ships as an MCP server so Claude can manage its own config programmatically

Under the Hood

Pure ES modules, zero npm dependencies (only MCP SDK for tool mode). 324 automated tests (124 unit + 200 Playwright E2E). Full CI/CD pipeline. Vanilla JS — no React, no build step, no bundler.

The effective resolution logic is extracted into a shared module (effective.mjs) used by both the dashboard and unit tests — same code path, no re-implementation.

Try It

npx @mcpware/claude-code-organizer

Zero config. Works on Linux, macOS, Windows, WSL.

First run auto-installs a /cco skill — after that, just type /cco in any Claude Code session to reopen.

mcpware / claude-code-organizer

Dashboard to manage Claude Code memories, configs, and MCP servers — security scanner for tool poisoning, context token budget tracker, duplicate cleanup, scope management. npx @mcpware/claude-code-organizer

Claude Code Organizer

AI agents: read AI_INDEX.md first. It is the navigation manifest for this codebase — where to find every module, how they connect, and where to look before making any claim about the code.

Claude Code Organizer (CCO) is a free, open-source dashboard that lets you manage all Claude Code configuration — memories, skills, MCP servers, settings, agents, rules, and hooks — across global and project scopes. It includes a security scanner for MCP tool poisoning and prompt injection, a per-item context token budget tracker, per-project MCP enable/disable controls, and bulk cleanup for duplicate configs. All without leaving the window.